A modeling method for dynamic intelligent prediction system of big data air conditioning system

Abstract

A modeling method for a dynamic intelligent prediction system of a big data air conditioning system includes the following steps: S1, establishing a BP neural network energy consumption prediction model for a cold source system: performing detrending and deperiodic processing on the known energy consumption data of the actual operation of the air conditioner: obtaining the energy consumption value of the time series after detrending and deperiodic processing; then establishing a three-layer BP neural network model and training the model, and finally converting the output prediction result of the BP network into the energy consumption demand prediction value of the cold source system of the air conditioning system; S2, optimizing the neural network with a genetic algorithm: encoding the weights and thresholds of the initial BP network into chromosomes, configuring the BP neural network through a genetic algorithm, and finally obtaining the optimized BP neural network. This design can not only effectively improve the prediction accuracy, but also achieve global optimization to make the prediction results more realistic.

Description

Technical Field

[0001] The present invention relates to a modeling method for a dynamic intelligent prediction system of a big data air-conditioning system, which is specifically suitable for reducing the relative error of training accuracy and prediction accuracy. Background Art

[0002] Domestic and foreign scholars have conducted many studies in the field of building energy conservation. At present, there are two main aspects to promote energy conservation and consumption reduction in buildings: first, accelerate the industrialization process of new buildings represented by prefabricated buildings. For example, the proportion of building energy consumption lost through doors and windows is as high as 51%, so the research on energy-saving doors and windows can be strengthened in terms of structural forms, materials, and assembly processes; second, reduce the energy consumption of building operation by improving the construction and management level of green buildings, promoting the high-quality application of renewable energy, improving the energy-saving standards of new buildings, and improving the energy efficiency of existing buildings. In terms of building operation energy consumption, the operation of the air-conditioning system is the main equipment for building energy consumption, which can account for 30% of the building operation energy consumption. Combining this field with sensor technology, communication technology and advanced control algorithms to study the optimization strategy of air-conditioning operation control in order to optimize energy utilization and achieve energy saving has always been the focus of research by domestic and foreign scholars. The accurate prediction of the energy consumption of the cold source system of the air-conditioning system is an important prerequisite for the energy-saving operation of the air-conditioning system.

[0003] Generally speaking, each building or household consists of multiple electrical loads that are used continuously or regularly, including lighting, motors, air conditioning systems, and other devices (such as computers, printers, TVs, mobile phones, etc.). But in general, air conditioning systems consume the most electrical energy. Due to the complexity of the internal systems of buildings, the consumption of energy and electricity in a specific building is usually affected by many factors, making it difficult to achieve optimal control. During the period of air conditioning power use, in order to obtain a stable energy supply, it has long relied on the experience and intuition of operators. However, in addition to a stable energy supply, consumers also require energy and cost savings. In addition, considering the energy demand that changes with seasons and time and the corresponding energy consumption costs, it is difficult for operators to operate equipment well.

[0004] Therefore, under the conditions of advocating energy conservation and emission reduction and consumers' demand for cost reduction, a consulting system or equipment that can provide energy consumption information to operators is needed. The main function of this consulting system is to accurately predict energy demand as the basic condition for operation planning, and to give appropriate system operation planning according to the predicted energy demand from the perspectives of energy saving, cost saving and ensuring stable energy supply. Summary of the invention

[0005] The purpose of the present invention is to overcome the problem of being easily trapped in local optimality in the prior art, and to provide a dynamic intelligent prediction system modeling method for a big data air-conditioning system that can take the overall situation into consideration.

[0006] To achieve the above objectives, the technical solution of the present invention is:

[0007] A big data air conditioning system dynamic intelligent prediction system modeling method, comprising the following steps:

[0008] S1. Establish a BP neural network energy consumption prediction model for the cold source system:

[0009] The known energy consumption data of the actual operation of the air conditioner is processed to remove the trend and period: the actual operation energy consumption data of the air conditioner is sampled at Δt = 1h to obtain the energy consumption value O of the time series. t (t=1,2,…), then the energy consumption data of the time series is detrended and deperiodic to obtain the energy consumption value w of the time series after detrended and deperiodic processing t ; Then a three-layer BP neural network model is established and the known air-conditioning operation data is used to train the model. Finally, the output prediction results of the BP network are converted into the energy consumption demand prediction value of the cold source system of the air-conditioning system;

[0010] S2. Genetic algorithm optimization of neural network:

[0011] The weights and thresholds of the initial BP network are encoded into chromosomes. After the initial population is generated, the next generation of new population is generated through selection, crossover and mutation operations among individuals in each generation of the population. The BP neural network is configured according to the gene value of the individual with the highest fitness value in the population after the iteration is completed, and the optimized BP neural network is obtained after training.

[0012] S1, establishing a BP neural network energy consumption prediction model for the cold source system:

[0013] S1.1. Detrend and de-periodize the known energy consumption data of the actual operation of the air conditioner:

[0014] The actual operating energy consumption data of the air conditioner is sampled at Δt = 1h to obtain the energy consumption value O of the time series t (t=1,2,…), then the energy consumption data of the time series is detrended and deperiodic to obtain the energy consumption value w of the time series after detrended and deperiodic processing t ;

[0015] S1.1.1. Detrending:

[0016] Removing trends from data helps to focus on analyzing the fluctuations of the data trend itself. The difference method is used to remove the linear trend in the data; the first-order difference is shown in formula 1:

[0017]

[0018] In the above formula, is a general difference operator, O t is the energy consumption of air conditioning at time t, O t-1 is the energy consumption value of the air conditioner at time t-1;

[0019] Perform the above operation d times, and get u by d-order difference. t , as shown in Formula 2:

[0020]

[0021] In the above formula, represents the number of combinations of selecting e elements from d elements, u t Represents the energy consumption value of the time series after detrending;

[0022] S1.1.2, De-cycle processing:

[0023] Since the energy demand changes in a 24-hour cycle, seasonal differences are used to remove the periodic trend of the time series. The de-periodic process is shown in Equation 3:

[0024]

[0025] in, is the de-periodicization operator, u t It represents the energy consumption value of the time series after detrending, s is the number of sampling times in a day, s=24;

[0026] Repeat the above operation D times, that is, after D de-periodic operations, we get w t , w t The value of is shown in formula 4:

[0027]

[0028] In the above formula, represents the number of combinations of selecting e elements from d elements, u t represents the energy consumption value of the time series after detrending, w t It represents the energy consumption value of the time series after detrending and deperiodic processing;

[0029] S1.2. Establish a three-layer neural network model

[0030] The BP neural network consists of n input layer neurons, h hidden layer neurons and m output layer neurons. Since there is only one output, the energy consumption prediction value, the number of neurons in the output layer is set to m = 1;

[0031] The input of the BP neural network includes: p time series data w after detrending and deperiodic processing t+l-1 、w t+l-2 ,…,w t+l-p and the corresponding outdoor temperature and relative humidity data; the number of neurons in the input layer of the BP neural network should be p+2, and the output of the network is: the energy consumption demand w predicted at time t+l (l=1,2,…,L) at time t t+l|t , where l≤s;

[0032] Assume that the input pattern of the network is x=(x1,x2,…,x n ) T , the output of the hidden layer is y=(y1,y2,…,y h ) T , the output of the output layer is z=(z1,z2,…,z m ) T , the transfer function from the hidden layer to the output layer is f, and the transfer function of the output layer is g, so we can get the input-output relationship between the hidden layer and the output layer and the input-output relationship of the output layer, as shown in Equation 5 and Equation 6:

[0033]

[0034]

[0035] In the above formula, w ij is the weight between the i-th neuron in the input layer and the j-th neuron in the hidden layer, η j is the threshold of the jth neuron in the hidden layer, w jk is the weight between the jth neuron in the hidden layer and the kth neuron in the output layer, θ k is the threshold of the kth neuron in the output layer;

[0036] The transfer function between the hidden layer and the output layer is f and the transfer function of the output layer is g. Both use the bipolar Sigmoid function as the activation function. The bipolar Sigmoid function f(x) is formula 7:

[0037]

[0038] In the formula, a is a constant between 0 and 10, and x represents the input amount of the input layer;

[0039] When selecting the optimal number of hidden layer neurons, the following formula is mainly used as a reference:

[0040]

[0041] Where a is a constant between 0 and 10, n represents the number of neurons in the input layer, and m represents the number of neurons in the output layer. The optimal number is determined by testing the network accuracy with different numbers of neurons in the hidden layer.

[0042] The reference for model training is the difference between the actual output and the target output. The objective function ε is shown in Equation 9:

[0043]

[0044] In the formula, w t|t-1 is the network predicted output value at time t, w t is the expected network output value, T represents the number of samples;

[0045] Select l r =0.01 as the learning rate of the BP neural network model;

[0046] The network output value w is obtained from formula 2 t+l|t With y t+l|t The relationship between them is as follows:

[0047]

[0048] From formula 4, we can get y t+l|t and the predicted value of air conditioning energy consumption z t+l|t The relationship between them is as follows:

[0049]

[0050] After the above two steps, the output value w of the BP neural network can be t+l|t Converted into the predicted value of air conditioning energy consumption demand z t+l|t .

[0051] S2, genetic algorithm optimization neural network:

[0052] S2.1. Construction of genetic algorithm optimization model:

[0053] S2.1.1, chromosome construction: for optimizing the BP neural network model, real number coding is used for chromosome coding, that is, individuals are represented as real number strings, including the weights between the input layer and the hidden layer, between the hidden layer and the output layer, and the thresholds of the hidden layer and the output layer; the result of modeling the data using only the BP neural network shows that the weights and thresholds of the constructed model are mostly concentrated between [-20, 20], so the upper limit of the gene value is set to 20 and the lower limit is -20; all the weights and thresholds of the network model are expressed by the individual gene values. After determining the network structure to be established, a BP neural network model with determined weights, thresholds and network structure can be obtained;

[0054] S2.1.2. Population initialization: Use linear interpolation to initialize individuals; the calculation formula is as follows:

[0055] a=lb+(ub-lb)*pick (12)

[0056] In the above formula, lb is the lower bound of the gene value, ub is the upper bound of the gene value, a represents the population initialization value, and pick is a random number in the interval [0, 1];

[0057] At the same time, a test function is constructed to test the feasibility of the chromosome to ensure that all genes are within their domains;

[0058] S2.1.3, determine the fitness function: construct a BP neural network model based on the individual gene value, then train the model with training data and perform data prediction output. The individual fitness value F is set to the sum of the absolute values ​​of the errors between the actual output of the model and the expected output. The calculation formula is as follows:

[0059]

[0060] Where m is the number of network output nodes m=1, z i|i-1 is the actual output of the ith node of the BP neural network, z i is the expected output of the i-th node, k is the coefficient;

[0061] S2.1.4. Population size:

[0062] The final effect of genetic iteration and the efficiency of genetic algorithm execution are both affected by the size of the population. p Take 100;

[0063] S2.1.5. Select operation:

[0064] The roulette method is used as the selection basis, and the fitness of individual i is recorded as f i , the probability of being selected is p i , then their calculation formulas are as follows:

[0065] f i = k / F i (14)

[0066]

[0067] In the above formula, F i represents the fitness value of the individual, k is the coefficient, N p is the number of individuals in the population N p =100;

[0068] Then generate a random number between 0 and 1 and compare it with the probability of the individual being selected calculated by formula 15. If the random number is greater than the probability of the individual being selected, the individual will be inherited to the next generation.

[0069] S2.1.6, Crossover probability and crossover operation:

[0070] The crossover probability is selected as 0.60, and the crossover probability P c It determines the frequency of the crossover operator. For each selected individual, a random number between 0 and 1 is generated to compare with the crossover probability. If it is less than the crossover probability, one or a gene on the individual is randomly selected to cross with another individual. Because the chromosome is constructed by real number coding, the real number crossover method is used as the crossover operation method, that is, the kth chromosome a k and chromosome a l Crossover is performed at position j, as follows:

[0071]

[0072] In the above formula, b is a random number between [0,1], but it is different from the random number used to determine whether to perform a crossover operation. The feasibility of the newly generated crossover individual is then tested. If feasible, it is inherited to the next generation population;

[0073] S2.1.7, Mutation probability and mutation operation:

[0074] The selected mutation probability P m is 0.1. For each selected individual, a random number between 0 and 1 is generated. If it is less than the mutation probability, a gene on the individual is randomly selected to perform the mutation operation. The jth gene a of the i-th individual is selected. ij To perform mutation, the operation method is as follows:

[0075]

[0076] Where, f(g) = r2(1-g / G) 2 , r2 is a random number, g represents the current evolutionary generation, G represents the termination evolutionary generation, r is a random number between [0,1], but it is different from the random number used to determine whether to perform the mutation operation. Then, the feasibility of the newly generated mutant individual is tested. If feasible, it will be inherited to the next generation population;

[0077] S2.1.8. Termination evolution algebra of genetic operation:

[0078] When the genetic algorithm evolutionary number is equal to the set termination evolutionary number G, the genetic algorithm operation ends.

[0079] Compared with the prior art, the present invention has the following beneficial effects:

[0080] 1. In the modeling method of a big data air-conditioning system dynamic intelligent prediction system of the present invention, the actual energy consumption data and meteorological data of the air-conditioning operation are firstly processed to remove trends and cycles. Secondly, the processed data is used as the main input, and the predicted energy consumption at the next moment is used as the output to establish a BP neural network model for energy consumption prediction. Then, the weights and thresholds of the model are encoded into chromosomes, the fitness function is determined, and the genetic algorithm is used to optimize it. The relative error of the training accuracy of the optimized model is reduced from 8.13% to 6.23%, and the relative error of the prediction accuracy is reduced from 8.69% to 6.78%. Accurate prediction of the energy consumption of the air-conditioning system is achieved, which provides a basis for studying the optimization control strategy of the air-conditioning system and will have a positive effect on the technological progress in the field of energy saving of the air-conditioning system. Therefore, this design can effectively improve the prediction accuracy.

[0081] 2. A BP neural network model for predicting the energy consumption of an air conditioning cold source system is proposed in a modeling method for a big data air conditioning system dynamic intelligent prediction system of the present invention. First, the data input into the BP network is detrended and deperiodic, then a three-layer BP neural network is established and trained, and finally the output prediction result of the BP network is converted into the energy consumption demand prediction value of the cold source system of the air conditioning system, which can effectively improve the accuracy of the training. Therefore, this design can effectively improve the accuracy of the training.

[0082] 3. In a modeling method of a dynamic intelligent prediction system for a big data air-conditioning system of the present invention, the weights and thresholds of the initial BP network are encoded as chromosomes. After the initial population is generated, the next generation of new populations is generated through selection, crossover and mutation operations between individuals in each generation of populations. According to the gene value of the individual with the highest fitness value in the population after the iteration is completed, the BP neural network is configured, and the optimized BP neural network is obtained after training. Since the genetic algorithm can perform global optimization, the prediction results of the optimized BP neural network are more accurate and comprehensive. Therefore, this design can achieve global optimization and make the prediction results more realistic. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 It is the operation flow chart of the BP algorithm of the present invention.

[0084] Figure 2 This is a diagram of an artificial neuron model.

[0085] Figure 3 This is a three-layer typical BP neural network model diagram.

[0086] Figure 4 This is the flow chart of genetic algorithm optimizing BP neural network

[0087] Figure 5 This is the training effect diagram of BP neural network without genetic algorithm optimization.

[0088] Figure 6 This is a diagram of the BP neural network training effect optimized using genetic algorithm.

[0089] Figure 7 This is the prediction effect diagram of the BP neural network without genetic algorithm optimization.

[0090] Figure 8 This is a graph of the prediction effect of the BP neural network optimized using a genetic algorithm. DETAILED DESCRIPTION

[0091] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0092] See also Figures 1 to 4 , a big data air conditioning system dynamic intelligent prediction system modeling method, comprising the following steps:

[0093] S1. Establish a BP neural network energy consumption prediction model for the cold source system:

[0094] The known energy consumption data of the actual operation of the air conditioner is processed to remove the trend and period: the actual operation energy consumption data of the air conditioner is sampled at Δt = 1h to obtain the energy consumption value O of the time series. t (t=1,2,…), then the energy consumption data of the time series is detrended and deperiodic to obtain the energy consumption value w of the time series after detrended and deperiodic processing t ; Then a three-layer BP neural network model is established and the known air-conditioning operation data is used to train the model. Finally, the output prediction results of the BP network are converted into the energy consumption demand prediction value of the cold source system of the air-conditioning system;

[0095] S2. Genetic algorithm optimization of neural network:

[0096] The weights and thresholds of the initial BP network are encoded into chromosomes. After the initial population is generated, the next generation of new population is generated through selection, crossover and mutation operations among individuals in each generation of the population. The BP neural network is configured according to the gene value of the individual with the highest fitness value in the population after the iteration is completed, and the optimized BP neural network is obtained after training.

[0097] S1, establishing a BP neural network energy consumption prediction model for the cold source system:

[0098] S1.1. Detrend and de-periodize the known energy consumption data of the actual operation of the air conditioner:

[0099] The actual operating energy consumption data of the air conditioner is sampled at Δt = 1h to obtain the energy consumption value O of the time series t(t=1,2,…), then the energy consumption data of the time series is detrended and deperiodic to obtain the energy consumption value w of the time series after detrended and deperiodic processing t ;

[0100] S1.1.1. Detrending:

[0101] Removing trends from data helps to focus on analyzing the fluctuations of the data trend itself. The difference method is used to remove the linear trend in the data; the first-order difference is shown in formula 1:

[0102]

[0103] In the above formula, is a general difference operator, O t is the energy consumption of air conditioning at time t, O t-1 is the energy consumption value of the air conditioner at time t-1;

[0104] Perform the above operation d times, and get u by d-order difference. t , as shown in Formula 2:

[0105]

[0106] In the above formula, represents the number of combinations of selecting e elements from d elements, u t Represents the energy consumption value of the time series after detrending;

[0107] S1.1.2, De-cycle processing:

[0108] Since the energy demand changes in a 24-hour cycle, seasonal differences are used to remove the periodic trend of the time series. The de-periodic process is shown in Equation 3:

[0109]

[0110] in, is the de-periodicization operator, u t It represents the energy consumption value of the time series after detrending, s is the number of sampling times in a day, s=24;

[0111] Repeat the above operation D times, that is, after D de-periodic operations, we get w t , w t The value of is shown in formula 4:

[0112]

[0113] In the above formula, represents the number of combinations of selecting e elements from d elements, u t represents the energy consumption value of the time series after detrending, wt It represents the energy consumption value of the time series after detrending and deperiodic processing;

[0114] S1.2. Establish a three-layer neural network model

[0115] The BP neural network consists of n input layer neurons, h hidden layer neurons and m output layer neurons. Since there is only one output, the energy consumption prediction value, the number of neurons in the output layer is set to m = 1;

[0116] The input of the BP neural network includes: p time series data w after detrending and deperiodic processing t+l-1 、w t+l-2 ,…,w t+l-p and the corresponding outdoor temperature and relative humidity data; the number of neurons in the input layer of the BP neural network should be p+2, and the output of the network is: the energy consumption demand w predicted at time t+l (l=1,2,…,L) at time t t+l|t , where l≤s;

[0117] Assume that the input pattern of the network is x=(x1,x2,…,x n ) T , the output of the hidden layer is y=(y1,y2,…,y h ) T , the output of the output layer is z=(z1,z2,…,z m ) T , the transfer function from the hidden layer to the output layer is f, and the transfer function of the output layer is g, so we can get the input-output relationship between the hidden layer and the output layer and the input-output relationship of the output layer, as shown in Equation 5 and Equation 6:

[0118]

[0119]

[0120] In the above formula, w ij is the weight between the i-th neuron in the input layer and the j-th neuron in the hidden layer, η j is the threshold of the jth neuron in the hidden layer, w jk is the weight between the jth neuron in the hidden layer and the kth neuron in the output layer, θ k is the threshold of the kth neuron in the output layer;

[0121] The transfer function between the hidden layer and the output layer is f and the transfer function of the output layer is g. Both use the bipolar Sigmoid function as the activation function. The bipolar Sigmoid function f(x) is formula 7:

[0122]

[0123] In the formula, a is a constant between 0 and 10, and x represents the input amount of the input layer;

[0124] When selecting the optimal number of hidden layer neurons, the following formula is mainly used as a reference:

[0125]

[0126] Where a is a constant between 0 and 10, n represents the number of neurons in the input layer, and m represents the number of neurons in the output layer. The optimal number is determined by testing the network accuracy with different numbers of neurons in the hidden layer.

[0127] The reference for model training is the difference between the actual output and the target output. The objective function ε is shown in Equation 9:

[0128]

[0129] In the formula, w t|t-1 is the network predicted output value at time t, w t is the expected network output value, T represents the number of samples;

[0130] Select l r =0.01 as the learning rate of the BP neural network model;

[0131] The network output value w is obtained from formula 2 t+l|t With y t+l|t The relationship between them is as follows:

[0132]

[0133] From formula 4, we can get y t+l|t and the predicted value of air conditioning energy consumption z t+l|t The relationship between them is as follows:

[0134]

[0135] After the above two steps, the output value w of the BP neural network can be t+l|t Converted into the predicted value of air conditioning energy consumption demand z t+l|t .

[0136] S2, genetic algorithm optimization neural network:

[0137] S2.1. Construction of genetic algorithm optimization model:

[0138] S2.1.1, chromosome construction: for optimizing the BP neural network model, real number coding is used for chromosome coding, that is, individuals are represented as real number strings, including the weights between the input layer and the hidden layer, between the hidden layer and the output layer, and the thresholds of the hidden layer and the output layer; the result of modeling the data using only the BP neural network shows that the weights and thresholds of the constructed model are mostly concentrated between [-20, 20], so the upper limit of the gene value is set to 20 and the lower limit is -20; all the weights and thresholds of the network model are expressed by the individual gene values. After determining the network structure to be established, a BP neural network model with determined weights, thresholds and network structure can be obtained;

[0139] S2.1.2. Population initialization: Use linear interpolation to initialize individuals; the calculation formula is as follows:

[0140] a=lb+(ub-lb)*pick (12)

[0141] In the above formula, lb is the lower bound of the gene value, ub is the upper bound of the gene value, a represents the population initialization value, and pick is a random number in the interval [0, 1];

[0142] At the same time, a test function is constructed to test the feasibility of the chromosome to ensure that all genes are within their domains;

[0143] S2.1.3, determine the fitness function: construct a BP neural network model based on the individual gene value, then train the model with training data and perform data prediction output. The individual fitness value F is set to the sum of the absolute values ​​of the errors between the actual output of the model and the expected output. The calculation formula is as follows:

[0144]

[0145] Where m is the number of network output nodes m=1, z i|i-1 is the actual output of the ith node of the BP neural network, z i is the expected output of the i-th node, k is the coefficient;

[0146] S2.1.4. Population size:

[0147] The final effect of genetic iteration and the efficiency of genetic algorithm execution are both affected by the size of the population. p Take 100;

[0148] S2.1.5. Select operation:

[0149] The roulette method is used as the selection basis, and the fitness of individual i is recorded as f i , the probability of being selected is p i , then their calculation formulas are as follows:

[0150] f i = k / F i (14)

[0151]

[0152] In the above formula, F i represents the fitness value of the individual, k is the coefficient, N p is the number of individuals in the population N p =100;

[0153] Then generate a random number between 0 and 1 and compare it with the probability of the individual being selected calculated by formula 15. If the random number is greater than the probability of the individual being selected, the individual will be inherited to the next generation.

[0154] S2.1.6, Crossover probability and crossover operation:

[0155] The crossover probability is selected as 0.60, and the crossover probability P c It determines the frequency of the crossover operator. For each selected individual, a random number between 0 and 1 is generated to compare with the crossover probability. If it is less than the crossover probability, one or a gene on the individual is randomly selected to cross with another individual. Because the chromosome is constructed by real number coding, the real number crossover method is used as the crossover operation method, that is, the kth chromosome a k and chromosome a l Crossover is performed at position j, as follows:

[0156]

[0157] In the above formula, b is a random number between [0,1], but it is different from the random number used to determine whether to perform a crossover operation. The feasibility of the newly generated crossover individual is then tested. If feasible, it is inherited to the next generation population;

[0158] S2.1.7, Mutation probability and mutation operation:

[0159] The selected mutation probability P m is 0.1. For each selected individual, a random number between 0 and 1 is generated. If it is less than the mutation probability, a gene on the individual is randomly selected to perform the mutation operation. The jth gene a of the i-th individual is selected. ij To perform mutation, the operation method is as follows:

[0160]

[0161] Where, f(g) = r2(1-g / G) 2, r2 is a random number, g represents the current evolutionary generation, G represents the termination evolutionary generation, r is a random number between [0,1], but it is different from the random number used to determine whether to perform the mutation operation. Then, the feasibility of the newly generated mutant individual is tested. If feasible, it will be inherited to the next generation population;

[0162] S2.1.8. Termination evolution algebra of genetic operation:

[0163] When the genetic algorithm evolutionary number is equal to the set termination evolutionary number G, the genetic algorithm operation ends.

[0164] The principle of the present invention is described as follows:

[0165] The establishment of the BP neural network model for energy consumption prediction of the air conditioning cold source system in S1. Since the air conditioning system is a complex system with multiple factors working together, the energy consumption load of the cold source system of the air conditioning system is affected by many factors inside and outside the system. Because the nonlinear mapping ability of the BP neural network is very strong, it can achieve the approximation of any nonlinear function. In order to make an accurate prediction of the energy consumption demand of the cold source system, this chapter proposes a BP neural network model for energy consumption prediction of the air conditioning cold source system. First, the data input to the BP network is detrended and de-periodic, and then a three-layer BP neural network is established and trained. Finally, the output prediction results of the BP network are converted into the energy consumption demand prediction value of the cold source system of the air conditioning system. In addition, a brief introduction to the operation process of the BP algorithm is given.

[0166] In S2, the genetic algorithm optimizes the BP neural network model. Since the BP neural network model is prone to fall into the local optimum when predicting the load of the air-conditioning cold source system, and the prediction effect is generally not very good, this chapter proposes to use the genetic algorithm to optimize the BP neural network model. The weights and thresholds of the initial BP network are encoded as chromosomes. After the initial population is generated, the next generation of new populations is generated through selection, crossover and mutation operations between individuals in each generation of the population. According to the gene value of the individual with the highest fitness value in the population after the iteration is completed, the BP neural network is configured, and the optimized BP neural network is obtained after training. Since the genetic algorithm can perform global optimization, the prediction effect of the optimized BP neural network is expected to be better.

[0167] Embodiment 1:

[0168] A big data air conditioning system dynamic intelligent prediction system modeling method, comprising the following steps:

[0169] S1. Establish a BP neural network energy consumption prediction model for the cold source system:

[0170] The known energy consumption data of the actual operation of the air conditioner is processed to remove the trend and period: the actual operation energy consumption data of the air conditioner is sampled at Δt = 1h to obtain the energy consumption value O of the time series.t (t=1,2,…), then the energy consumption data of the time series is detrended and deperiodic to obtain the energy consumption value w of the time series after detrended and deperiodic processing t ; Then a three-layer BP neural network model is established and the known air-conditioning operation data is used to train the model. Finally, the output prediction results of the BP network are converted into the energy consumption demand prediction value of the cold source system of the air-conditioning system;

[0171] S2. Genetic algorithm optimization of neural network:

[0172] The weights and thresholds of the initial BP network are encoded into chromosomes. After the initial population is generated, the next generation of new population is generated through selection, crossover and mutation operations among individuals in each generation of the population. The BP neural network is configured according to the gene value of the individual with the highest fitness value in the population after the iteration is completed, and the optimized BP neural network is obtained after training.

[0173] Embodiment 2:

[0174] Embodiment 2 is substantially the same as Embodiment 1, except that:

[0175] S1, establishing a BP neural network energy consumption prediction model for the cold source system:

[0176] S1.1. Detrend and de-periodize the known energy consumption data of the actual operation of the air conditioner:

[0177] The actual operating energy consumption data of the air conditioner is sampled at Δt = 1h to obtain the energy consumption value O of the time series t (t=1,2,…), then the energy consumption data of the time series is detrended and deperiodic to obtain the energy consumption value w of the time series after detrended and deperiodic processing t ;

[0178] S1.1.1. Detrending:

[0179] Removing trends from data helps to focus on analyzing the fluctuations of the data trend itself. The difference method is used to remove the linear trend in the data; the first-order difference is shown in formula 1:

[0180]

[0181] In the above formula, is a general difference operator, O t is the energy consumption of air conditioning at time t, O t-1 is the energy consumption value of the air conditioner at time t-1;

[0182] Perform the above operation d times, and get u by d-order difference. t , as shown in Formula 2:

[0183]

[0184] In the above formula, represents the number of combinations of selecting e elements from d elements, u t Represents the energy consumption value of the time series after detrending;

[0185] S1.1.2, De-cycle processing:

[0186] Since the energy demand changes in a 24-hour cycle, seasonal differences are used to remove the periodic trend of the time series. The de-periodic process is shown in Equation 3:

[0187]

[0188] in, is the de-periodicization operator, u t It represents the energy consumption value of the time series after detrending, s is the number of sampling times in a day, s=24;

[0189] Repeat the above operation D times, that is, after D de-periodic operations, we get w t , w t The value of is shown in formula 4:

[0190]

[0191] In the above formula, represents the number of combinations of selecting e elements from d elements, u t represents the energy consumption value of the time series after detrending, w t It represents the energy consumption value of the time series after detrending and deperiodic processing;

[0192] S1.2. Establish a three-layer neural network model

[0193] The BP neural network consists of n input layer neurons, h hidden layer neurons and m output layer neurons. Since there is only one output, the energy consumption prediction value, the number of neurons in the output layer is set to m = 1;

[0194] The input of the BP neural network includes: p time series data w after detrending and deperiodic processing t+l-1 、w t+l-2 ,…,w t+l-p and the corresponding outdoor temperature and relative humidity data; the number of neurons in the input layer of the BP neural network should be p+2, and the output of the network is: the energy consumption demand w predicted at time t+l (l=1,2,…,L) at time t t+l|t , where l≤s;

[0195] Assume that the input pattern of the network is x=(x1,x2,…,x n ) T , the output of the hidden layer is y=(y1,y2,…,y h ) T , the output of the output layer is z=(z1,z2,…,z m ) T , the transfer function from the hidden layer to the output layer is f, and the transfer function of the output layer is g, so we can get the input-output relationship between the hidden layer and the output layer and the input-output relationship of the output layer, as shown in Equation 5 and Equation 6:

[0196]

[0197]

[0198] In the above formula, w ij is the weight between the i-th neuron in the input layer and the j-th neuron in the hidden layer, η j is the threshold of the jth neuron in the hidden layer, w jk is the weight between the jth neuron in the hidden layer and the kth neuron in the output layer, θ k is the threshold of the kth neuron in the output layer;

[0199] The transfer function between the hidden layer and the output layer is f and the transfer function of the output layer is g. Both use the bipolar Sigmoid function as the activation function. The bipolar Sigmoid function f(x) is formula 7:

[0200]

[0201] In the formula, a is a constant between 0 and 10, and x represents the input amount of the input layer;

[0202] When selecting the optimal number of hidden layer neurons, the following formula is mainly used as a reference:

[0203]

[0204] Where a is a constant between 0 and 10, n represents the number of neurons in the input layer, and m represents the number of neurons in the output layer. The optimal number is determined by testing the network accuracy with different numbers of neurons in the hidden layer.

[0205] The reference for model training is the difference between the actual output and the target output. The objective function ε is shown in Equation 9:

[0206]

[0207] In the formula, w t|t-1 is the network predicted output value at time t, w tis the expected network output value, T represents the number of samples;

[0208] Select l r =0.01 as the learning rate of the BP neural network model;

[0209] The network output value w is obtained from formula 2 t+l|t With y t+l|t The relationship between them is as follows:

[0210]

[0211] From formula 4, we can get y t+l|t and the predicted value of air conditioning energy consumption z t+l|t The relationship between them is as follows:

[0212]

[0213] After the above two steps, the output value w of the BP neural network can be t+l|t Converted into the predicted value of air conditioning energy consumption demand z t+l|t .

[0214] S2, genetic algorithm optimization neural network:

[0215] S2.1. Construction of genetic algorithm optimization model:

[0216] S2.1.1, chromosome construction: for optimizing the BP neural network model, real number coding is used for chromosome coding, that is, individuals are represented as real number strings, including the weights between the input layer and the hidden layer, between the hidden layer and the output layer, and the thresholds of the hidden layer and the output layer; the result of modeling the data using only the BP neural network shows that the weights and thresholds of the constructed model are mostly concentrated between [-20, 20], so the upper limit of the gene value is set to 20 and the lower limit is -20; all the weights and thresholds of the network model are expressed by the individual gene values. After determining the network structure to be established, a BP neural network model with determined weights, thresholds and network structure can be obtained;

[0217] S2.1.2. Population initialization: Use linear interpolation to initialize individuals; the calculation formula is as follows:

[0218] a=lb+(ub-lb)*pick (12)

[0219] In the above formula, lb is the lower bound of the gene value, ub is the upper bound of the gene value, a represents the population initialization value, and pick is a random number in the interval [0, 1];

[0220] At the same time, a test function is constructed to test the feasibility of the chromosome to ensure that all genes are within their domains;

[0221] S2.1.3, determine the fitness function: construct a BP neural network model based on the individual gene value, then train the model with training data and perform data prediction output. The individual fitness value F is set to the sum of the absolute values ​​of the errors between the actual output of the model and the expected output. The calculation formula is as follows:

[0222]

[0223] Where m is the number of network output nodes m=1, z i|i-1 is the actual output of the ith node of the BP neural network, z i is the expected output of the i-th node, k is the coefficient;

[0224] S2.1.4. Population size:

[0225] The final effect of genetic iteration and the efficiency of genetic algorithm execution are both affected by the size of the population. p Take 100;

[0226] S2.1.5. Select operation:

[0227] The roulette method is used as the selection basis, and the fitness of individual i is recorded as f i , the probability of being selected is p i , then their calculation formulas are as follows:

[0228] f i = k / F i (14)

[0229]

[0230] In the above formula, F i represents the fitness value of the individual, k is the coefficient, N p is the number of individuals in the population N p =100;

[0231] Then generate a random number between 0 and 1 and compare it with the probability of the individual being selected calculated by formula 15. If the random number is greater than the probability of the individual being selected, the individual will be inherited to the next generation.

[0232] S2.1.6, Crossover probability and crossover operation:

[0233] The crossover probability is selected as 0.60, and the crossover probability P cIt determines the frequency of the crossover operator. For each selected individual, a random number between 0 and 1 is generated to compare with the crossover probability. If it is less than the crossover probability, one or a gene on the individual is randomly selected to cross with another individual. Because the chromosome is constructed by real number coding, the real number crossover method is used as the crossover operation method, that is, the kth chromosome a k and chromosome a l Crossover is performed at position j, as follows:

[0234]

[0235] In the above formula, b is a random number between [0,1], but it is different from the random number used to determine whether to perform a crossover operation. The feasibility of the newly generated crossover individual is then tested. If feasible, it is inherited to the next generation population;

[0236] S2.1.7, Mutation probability and mutation operation:

[0237] The selected mutation probability P m is 0.1. For each selected individual, a random number between 0 and 1 is generated. If it is less than the mutation probability, a gene on the individual is randomly selected to perform the mutation operation. The jth gene a of the i-th individual is selected. ij To perform mutation, the operation method is as follows:

[0238]

[0239] Where, f(g) = r2(1-g / G) 2 , r2 is a random number, g represents the current evolutionary generation, G represents the termination evolutionary generation, r is a random number between [0,1], but it is different from the random number used to determine whether to perform the mutation operation. Then, the feasibility of the newly generated mutant individual is tested. If feasible, it will be inherited to the next generation population;

[0240] S2.1.8. Termination evolution algebra of genetic operation:

[0241] When the genetic algorithm evolutionary number is equal to the set termination evolutionary number G, the genetic algorithm operation ends.

[0242] Embodiment 3:

[0243] Embodiment 3 is substantially the same as Embodiment 2, except that:

[0244] The data sample comes from a family residence located in Urumqi, Xinjiang Uygur Autonomous Region, with a temperate continental climate. The data contains the minute-by-minute energy consumption data of the air conditioning power consumption system and meteorological data from 2006 to 2009. The meteorological data includes outdoor temperature and relative humidity. The data from 2006 to 2008 are taken as the training data of the neural network, and the data from May and June 2009 are taken as the test data of the neural network. The sampling time interval is set to Δt = 1h, and the hourly air conditioning energy consumption is summed to obtain the hourly air conditioning energy consumption data of the family. Finally, a total of 730×24=17520 training data and 61×24=1464 test data are obtained.

[0245] In order to study the influence of genetic algorithm and model parameters on BP neural network training effect and prediction results, this chapter mainly makes the following two comparisons:

[0246] (1) By comparing the training and prediction results of the BP neural network model before and after the genetic algorithm is used to optimize it, the necessity of optimization and the optimization effect of the genetic algorithm are verified. The selection of the relevant parameters of the model and algorithm are as described above. It should be pointed out that the optimized model is used in the study (2).

[0247] (2) The effect of the number of hidden layer neurons in the BP neural network on the model training effect and prediction effect was studied. The test range was h = 2 to 11, and the training effect and test effect when h = 5 were taken as the reference benchmark. The number of hidden layer neurons h = 5 was also used in the study of (1).

[0248] In all the above test studies, the order of detrending difference operation is set to d=1, the order of deperiodicity operation is set to D=1, the number of sampling times for prediction is set to L=1, and p=3 is set. In addition, the outdoor temperature and relative humidity of the air are added, so the number of neurons in the input layer is 5.

[0249] Optimization effect of genetic algorithm: First, the models constructed by BP neural network with and without genetic algorithm optimization are trained using training data respectively. Figure 5 and Figure 6 The training effects of the actual and predicted energy demand of the cold source system obtained by the BP network model without and with genetic algorithm optimization are compared. The predicted energy demand value is the output of the two models converted according to the conversion relationship mentioned above. Figure 5 It can be seen that the training effect of the BP neural network model without genetic algorithm optimization sometimes fluctuates greatly, and the error between the predicted and actual energy consumption demand sometimes becomes too large. This phenomenon mainly occurs when the energy consumption demand is small. Therefore, this result is not suitable for practical applications. On the other hand, according to Figure 6It can be seen that the prediction effect is greatly improved when the energy consumption demand is relatively small, and the prediction is consistent with the actual energy consumption demand.

[0250] Next, the predicted data are used to predict the energy consumption of the models constructed by the BP neural network without and with genetic algorithm optimization. Figure 7 and Figure 8 The prediction effects of the actual and predicted energy consumption demands of the cooling source system obtained without and with the BP network model optimized by genetic algorithm were compared.

[0251] Depend on Figure 7 and Figure 8 It can be seen that the models constructed by the two methods have similar effects in terms of data prediction and data training fitting. The predicted value of the model constructed by using genetic algorithm optimization is closer to the actual measured value.

[0252] Table 1 lists the absolute and relative values ​​of the average error in four cases. Each relative value is obtained by dividing the corresponding absolute value by the difference between the maximum and minimum values ​​of the actual value during model training and prediction training. For training and prediction, the absolute and relative values ​​corresponding to the BP neural network model optimized by the genetic algorithm are much smaller.

[0253] Table 1 Absolute and relative values ​​of the average error of the model optimized by genetic algorithm or not

[0254]

[0255] The above results show that the establishment of an energy consumption prediction model for the cold source system of an air-conditioning system requires a global optimization method, and the genetic algorithm can overcome the shortcomings of the BP neural network that is prone to fall into local optimality to a certain extent, and obtain a better model, thereby achieving better energy consumption prediction results.

[0256] Influence of the number of hidden layer neurons: In order to explore the influence of the number of hidden layer neurons on the constructed BP neural network model, this paper constructs models with different numbers of hidden layer neurons h, and measures the absolute error and relative error in model training and model prediction to illustrate its influence on the model. As mentioned above, the number of hidden layer neurons tested ranges from h = 2 to 10. Table 2 lists the absolute and relative values ​​of the average error in all cases.

[0257] Table 2 The influence of absolute and relative values ​​of the number of hidden layer neurons error

[0258]

[0259] As can be seen from the table, when the number of hidden layer neurons is too many or too few, the network model is more suitable for the network training data, while for the prediction data, the model effect becomes worse. When the number of hidden layer neurons is 5, the training accuracy and prediction accuracy of the BP neural network model are both high, so 5 is selected as the number of hidden layer neurons for building the model.

[0260] The above results show that the number of neurons in the hidden layer has a certain impact on the training and prediction effects of the network model. It is necessary to quantify these effects. Quantitative indicators can help determine the optimal number of neurons to achieve better training and prediction effects of the model.

Claims

1. A big data air conditioning system dynamic intelligent prediction system modeling method, characterized by: The intelligent prediction method comprises the following steps: S1. Establish a BP neural network energy consumption prediction model for the cold source system: S1.

1. Detrending and deperiodicating the known energy consumption data of the actual operation of the air conditioner: The actual energy consumption data of the air conditioner is sampled at Δt = 1h to obtain the energy consumption value O of the time series. t (t=1,2,…), then the energy consumption data of the time series is detrended and deperiodic to obtain the energy consumption value w of the time series after detrended and deperiodic processing t ; S1.1.

1. Detrending: Removing trends from data helps to focus on analyzing the fluctuations of the data trend itself. The difference method is used to remove the linear trend in the data; the first-order difference is shown in formula 1: In the above formula, is a general difference operator, O t is the energy consumption of air conditioning at time t, O t-1 is the energy consumption value of the air conditioner at time t-1; Perform the above operation d times, and get u by d-order difference. t , as shown in Formula 2: In the above formula, represents the number of combinations of selecting e elements from d elements, u t Represents the energy consumption value of the time series after detrending; S1.1.2, De-cycle processing: Since the energy demand changes in a 24-hour cycle, seasonal differences are used to remove the periodic trend of the time series. The de-periodic process is shown in Equation 3: in, is the de-periodicization operator, u t It represents the energy consumption value of the time series after detrending, s is the number of sampling times in a day, s=24; Repeat the above operation D times, that is, after D de-periodic operations, we get w t , w t The value of is shown in formula 4: In the above formula, represents the number of combinations of selecting e elements from d elements, u t represents the energy consumption value of the time series after detrending, w t It represents the energy consumption value of the time series after detrending and deperiodic processing; S1.

2. Establish a three-layer neural network model: A three-layer BP neural network model is established and the known air-conditioning operation data is used to train the model. Finally, the output prediction results of the BP network are converted into the energy consumption demand prediction value of the cold source system of the air-conditioning system. The BP neural network consists of n input layer neurons, h hidden layer neurons and m output layer neurons. Since there is only one output, the energy consumption prediction value, the number of neurons in the output layer is set to m = 1; The input of the BP neural network includes: p time series data w after detrending and deperiodic processing t+l-1 、w t+l-2 ,…,w t+l-p and the corresponding outdoor temperature and relative humidity data; the number of neurons in the input layer of the BP neural network should be p+2, and the output of the network is: the energy consumption demand w predicted at time t+l (l=1,2,...,L) at time t t+l|t , where l≤s; Assume that the input pattern of the network is x=(x1, x2,..., x n ) T , the output of the hidden layer is y = (y1, y2, ..., y h ) T , the output of the output layer is z=(z1,z2,...,z m ) T , the transfer function from the hidden layer to the output layer is f, and the transfer function of the output layer is gg, so we can get the input-output relationship between the hidden layer and the output layer and the input-output relationship of the output layer, as shown in Equation 5 and Equation 6: In the above formula, w ij is the weight between the i-th neuron in the input layer and the j-th neuron in the hidden layer, η j is the threshold of the jth neuron in the hidden layer, w jk is the weight between the jth neuron in the hidden layer and the kth neuron in the output layer, θ k is the threshold of the kth neuron in the output layer; The transfer function between the hidden layer and the output layer is f and the transfer function of the output layer is gg. Both use the bipolar Sigmoid function as the activation function. The bipolar Sigmoid function f(x) is formula 7: In the formula, a is a constant between 0 and 10, and x represents the input amount of the input layer; When selecting the optimal number of hidden layer neurons, the following formula is mainly used as a reference: Where a is a constant between 0 and 10, n represents the number of neurons in the input layer, and m represents the number of neurons in the output layer. The optimal number is determined by testing the network accuracy with different numbers of neurons in the hidden layer. The reference for model training is the difference between the actual output and the target output. The objective function ε is shown in Equation 9: In the formula, w t|t-1 is the network predicted output value at time t, w t is the expected network output value, T represents the number of samples; Select l r =0.01 as the learning rate of the BP neural network model; The network output value w is obtained from formula 2 t+l|t With y t+l|t The relationship between them is as follows: From formula 4, we can get y t+l|t and the predicted value of air conditioning energy consumption z t+l|t The relationship between them is as follows: After the above two steps, the output value w of the BP neural network can be t+l|t Converted into the predicted value of air conditioning energy consumption demand z t+l|t ; S2. Genetic algorithm optimization of neural network: The weights and thresholds of the initial BP network are encoded into chromosomes. After the initial population is generated, the next generation of new population is generated through selection, crossover and mutation operations among individuals in each generation of the population. The BP neural network is configured according to the gene value of the individual with the highest fitness value in the population after the iteration is completed, and the optimized BP neural network is obtained after training.

2. A big data air conditioning system dynamic intelligent prediction system modeling method according to claim 1, characterized in that: S2, genetic algorithm optimization neural network: S2.

1. Construction of genetic algorithm optimization model: S2.1.

1. Construct chromosomes: For optimizing the BP neural network model, real number coding is used for chromosome coding, that is, individuals are represented as real number strings, including the weights between the input layer and the hidden layer, between the hidden layer and the output layer, and the thresholds of the hidden layer and the output layer; The results of using only BP neural network to model the data show that the weights and thresholds of the constructed model are mostly concentrated between [-20, 20], so the upper limit of the gene value is set to 20 and the lower limit is -20; all the weights and thresholds of the network model are expressed by the individual gene values. After determining the network structure to be established, a BP neural network model with determined weights, thresholds and network structure can be obtained; S2.1.

2. Population initialization: Use linear interpolation to initialize individuals; the calculation formula is as follows: a=lb+(ub-lb)*pick(12) In the above formula, lb is the lower bound of the gene value, ub is the upper bound of the gene value, a represents the population initialization value, and pick is [0, 1] Random numbers in the interval; At the same time, a test function is constructed to test the feasibility of the chromosome to ensure that all genes are within their domains; S2.1.3, determine the fitness function: construct a BP neural network model based on the gene value of the individual, then train the model with training data and perform data prediction output. The individual fitness value F is set to the sum of the absolute values ​​of the errors between the actual output of the model and the expected output. The calculation formula is as follows: Where m is the number of network output nodes m=1, z i|i-1 is the actual output of the ith node of the BP neural network, z i is the expected output of the i-th node, k is the coefficient; S2.1.

4. Population size: The final effect of genetic iteration and the efficiency of genetic algorithm execution are both affected by the size of the population. p Take 100; S2.1.

5. Select operation: The roulette method is used as the selection basis, and the fitness of individual i is recorded as f i , the probability of being selected is p i , their calculation formulas are as follows: f i =k / F i (14) In the above formula, F i represents the fitness value of the individual, k is the coefficient, N p is the number of individuals in the population N p =100; Then generate a random number between 0 and 1 and compare it with the probability of the individual being selected calculated by formula 15. If the random number is greater than the probability of the individual being selected, the individual will be inherited to the next generation. S2.1.6, Crossover probability and crossover operation: The crossover probability is selected as 0.60, and the crossover probability P c It determines the frequency of the crossover operator. For each selected individual, a random number between 0 and 1 is generated to compare with the crossover probability. If it is less than the crossover probability, one or a gene on the individual is randomly selected to cross with another individual. Because the chromosomes are constructed using real number coding, the real number crossover method is used as the crossover operation method, that is, the kth chromosome a k and chromosome l l Crossover is performed at position j, as follows: In the above formula, b is a random number between [0,1], but it is different from the random number used to determine whether to perform a crossover operation. The feasibility of the newly generated crossover individual is then tested. If feasible, it is inherited to the next generation population; S2.1.7, Mutation probability and mutation operation: The selected mutation probability P m is 0.

1. For each selected individual, a random number between 0 and 1 is generated. If it is less than the mutation probability, a gene on the individual is randomly selected to perform the mutation operation. The jth gene a of the i-th individual is selected. ij To perform mutation, the operation method is as follows: Where, f(g) = r2(1-g / G) 2 , r2 is a random number, g represents the current evolutionary generation, G represents the termination evolutionary generation, r is a random number between [0,1], but it is different from the random number used to determine whether to perform the mutation operation. Then, the feasibility of the newly generated mutant individual is tested. If feasible, it will be inherited to the next generation population; S2.1.

8. Termination evolution algebra of genetic operation: When the genetic algorithm evolutionary number is equal to the set termination evolutionary number G, the genetic algorithm operation ends.

Images