Central air conditioning system indoor thermal environment prediction method based on data driving

Abstract

The invention discloses a central air-conditioning system indoor thermal environment prediction method based on data driving, which relates to the field of energy conservation and comprises the following steps: screening indoor thermal environment influence factor variables of a building, and performing data preprocessing on the thermal environment influence factor variables; optimizing the BP neural network by using an improved grey wolf optimization algorithm to obtain a data-driven central air-conditioning system indoor thermal environment prediction model; the preprocessed thermal environment influence factor variables are input into a data-driven central air conditioning system indoor thermal environment prediction model, and the central air conditioning system indoor thermal environment is predicted; and performing performance evaluation on the data-driven central air-conditioning system indoor thermal environment prediction model by using the evaluation index to complete data-driven central air-conditioning system indoor thermal environment prediction. The problems that an existing method is low in prediction precision, and a prediction model is prone to being affected by equipment operation data are solved.

Description

Technical Field

[0001] The present invention relates to the field of energy conservation, and in particular to a data-driven indoor thermal environment prediction method for a central air-conditioning system. Background Art

[0002] Central air conditioning is a basic equipment used to control and adjust the ambient temperature, relative humidity, air flow velocity, cleanliness and other parameters in the building. It plays a key role in the building and aims to provide people with a comfortable working and living environment. The energy consumption of the central air conditioning system is mainly composed of the energy consumption of the air conditioning cold and heat source and the transmission energy consumption of the air and water system of the air conditioning. At present, the energy-saving research on the central air conditioning system mainly focuses on two aspects. One is to save energy in the air conditioning cold and heat source, such as the use of new energy and some optimization of cooling and heating control. The other is to study the energy-saving of the air and water transmission and distribution system in the air conditioning device, such as variable flow, variable air volume, variable supply and return water temperature difference and other technical means. Traditional central air conditioning systems usually adopt a relatively single control strategy or rely on empirical rules for control, ignoring the impact of many changing factors such as indoor and outdoor environment, equipment characteristics, user behavior, etc. on the environment and energy consumption, and cannot accurately optimize the real-time energy demand and indoor comfort. In order to enable the central air conditioning system to accurately control the indoor thermal environment comfort of the building, it is necessary to first establish an accurate building thermal environment prediction model. The prediction performance of the model is crucial for the precise control of the thermal environment. By establishing an accurate building thermal environment prediction model, we can better understand the changes in the building's indoor thermal environment, thereby more effectively controlling the operation of the central air-conditioning system and improving the comfort of the building's indoor thermal environment.

[0003] However, there are many factors that affect the indoor thermal environment of buildings, including not only many indoor and outdoor environmental parameters, but also various factors such as the equipment operating parameters of the central air-conditioning system. These factors are interrelated and coupled, and together affect the thermal environment, so it is very difficult to establish an accurate mathematical mechanism model for them. In addition, the air-conditioning system will deviate from the design operating conditions for a long time in actual operation. If the actual thermal performance of each equipment parameter is analyzed and processed based on the factory data of the equipment, the prediction accuracy of the established thermal environment model will be greatly affected, making it difficult to use in actual control. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a data-driven central air-conditioning system indoor thermal environment prediction method which solves the problem that the existing method has low prediction accuracy and the prediction model is easily affected by equipment operation data.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a data-driven central air-conditioning system indoor thermal environment prediction method, comprising the following steps:

[0006] S1: Screen the variables of thermal environment influencing factors in the building interior and perform data preprocessing on the variables of thermal environment influencing factors;

[0007] S2: The BP neural network is optimized using the improved grey wolf optimization algorithm to obtain a data-driven indoor thermal environment prediction model for the central air-conditioning system;

[0008] S3: input the preprocessed thermal environment influencing factor variables into the data-driven central air-conditioning system indoor thermal environment prediction model to predict the indoor thermal environment of the central air-conditioning system;

[0009] S4: Use the evaluation index to evaluate the performance of the data-driven central air-conditioning system indoor thermal environment prediction model, and complete the data-driven central air-conditioning system indoor thermal environment prediction.

[0010] Furthermore, the variables of the factors affecting the indoor thermal environment of the building in S1 include outdoor meteorological data, indoor thermal disturbance data, central air-conditioning operation parameter data and indoor thermal environment data at the current moment;

[0011] The outdoor meteorological data includes outdoor temperature and outdoor humidity;

[0012] The indoor thermal disturbance data includes heat dissipation of personnel, lighting and office equipment;

[0013] The central air-conditioning operation parameter data includes the water supply temperature of the cold and hot source equipment, the water supply flow rate and air volume of the room terminal equipment;

[0014] The current indoor thermal environment data includes indoor temperature and indoor humidity.

[0015] Furthermore, the data preprocessing of the thermal environment influencing factor variables in S1 includes correlation analysis and data normalization;

[0016] The correlation analysis uses the Spearman correlation coefficient ρ s , the formula is:

[0017]

[0018] Among them, R(·) is the rank of the variable, x l and l are different variables, the subscript l represents the lth value, and N is the number of variables;

[0019] The data normalization adopts range normalization or standard deviation normalization.

[0020] Furthermore, the BP neural network is optimized using the improved grey wolf optimization algorithm in S2, including the following steps:

[0021] S21: Initialize the gray wolf population and set the number of iterations and fitness function;

[0022] S22: According to the set fitness function, the position of the gray wolf population is updated by adopting a dynamic weight update strategy based on the step-length Euclidean distance, and the fitness value of the updated gray wolf population is calculated;

[0023] S23: Determine whether the current iteration meets the set number of iterations. If not, return to step S22. If yes, obtain the optimal individual based on the updated fitness value of the gray wolf population, and use the optimal individual as the initial weight and threshold of the BP neural network for network training, and enter step S24;

[0024] S24: Calculate the output error of the BP neural network and adjust the initial weight and threshold through error back propagation;

[0025] S25: Based on the adjusted initial weights and thresholds, calculate the average error of the BP neural network at the current moment, and determine whether the current iteration meets the preset conditions. If not, return to step S24. If so, output the BP neural network prediction result to complete the optimization of the BP neural network.

[0026] Furthermore, in S21, the gray wolf population is initialized using Circle chaotic mapping, and the formula is:

[0027]

[0028] Among them, x m+1 is the chaotic value of the m+1th iteration, x m is the chaotic value of the mth iteration, and mod(·) is the modulo operation.

[0029] Furthermore, the dynamic weight update strategy based on the step length Euclidean distance in S22 is:

[0030]

[0031] Among them, ω 1 ,ω 2 and ω 3 For different weights, X 1 , X 2 and X 3 are different individuals, X(·) is the step length Euclidean distance, and t is the current iteration number.

[0032] Furthermore, in S24, the initial weight and threshold are adjusted by error back propagation, and the formula is:

[0033]

[0034] Among them, Wj is the weight corresponding to the jth input value, b j is the threshold corresponding to the jth input value, α is the learning rate, and E is the error criterion function that needs to be reduced.

[0035] Furthermore, the convergence factor r of the improved grey wolf optimization algorithm in S2 is:

[0036]

[0037] Among them, t is the current iteration number and T is the maximum iteration number.

[0038] Furthermore, the evaluation indicators in S4 include root mean square error, mean absolute percentage error and complex determination coefficient, and the formula is:

[0039]

[0040] Among them, RMSE is the root mean square error, n is the number of sample data, is the model prediction value of sample i, p i is the actual value of sample i;

[0041]

[0042] Among them, MAPE is the mean absolute percentage error;

[0043]

[0044] Among them, R 2 is the complex determination coefficient, is the average of the actual values ​​of sample i.

[0045] The beneficial effects of the present invention are as follows: the present invention combines the improved grey wolf algorithm with the BP neural network to construct a data-driven thermal environment prediction model. The data-driven model obtained by reasonable analysis based on real operating data can effectively solve the problem of model distortion and obtain an accurate prediction structure. By comparing different model algorithms, it is concluded that the improved GWO-BP neural network model has better prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of a data-driven method for predicting the indoor thermal environment of a central air-conditioning system.

[0047] Figure 2 The comparison of convergence factors before and after improvement is shown in the figure.

[0048] Figure 3 This is a schematic diagram of the central air-conditioning system experimental platform.

[0049] Figure 4 This is the fitness update curve of the gray wolf population in the room temperature prediction model.

[0050] Figure 5 Comparison chart of predicted value and actual value of room temperature in room 1.

[0051] Figure 6 This is a comparison chart of the predicted value and the actual value of the room temperature in room 2.

[0052] Figure 7 This is a comparison chart of the predicted value and the actual value of the room temperature in room 3. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0054] like Figure 1 As shown, a data-driven method for predicting indoor thermal environment of a central air-conditioning system includes the following steps:

[0055] S1: Screen the variables of thermal environment influencing factors in the building interior and perform data preprocessing on the variables of thermal environment influencing factors;

[0056] S2: The BP neural network is optimized using the improved grey wolf optimization algorithm to obtain a data-driven indoor thermal environment prediction model for the central air-conditioning system;

[0057] S3: input the preprocessed thermal environment influencing factor variables into the data-driven central air-conditioning system indoor thermal environment prediction model to predict the indoor thermal environment of the central air-conditioning system;

[0058] S4: Use the evaluation index to evaluate the performance of the data-driven central air-conditioning system indoor thermal environment prediction model, and complete the data-driven central air-conditioning system indoor thermal environment prediction.

[0059] The variables of the factors affecting the indoor thermal environment of the building in S1 include outdoor meteorological data, indoor thermal disturbance data, central air-conditioning operation parameter data and the indoor thermal environment data at the current moment;

[0060] The outdoor meteorological data includes outdoor temperature and outdoor humidity;

[0061] The indoor thermal disturbance data includes heat dissipation of personnel, lighting and office equipment;

[0062] The central air-conditioning operation parameter data includes the water supply temperature of the cold and hot source equipment, the water supply flow rate and air volume of the room terminal equipment;

[0063] The current indoor thermal environment data includes indoor temperature and indoor humidity.

[0064] In said S1, data preprocessing of the thermal environment influencing factor variables includes correlation analysis and data normalization;

[0065] The correlation analysis uses the Spearman correlation coefficient ρ s , the formula is:

[0066]

[0067] Among them, R(·) is the rank of the variable, x l and l are different variables, the subscript l represents the lth value, and N is the number of variables;

[0068] The data normalization adopts range normalization or standard deviation normalization.

[0069] The two commonly used correlation coefficients are the Pearson correlation coefficient and the Spearman correlation coefficient. The Spearman correlation coefficient calculates the degree of correlation between two variables based on the ranking of the two variables. Unlike the Pearson correlation analysis, the Spearman correlation analysis does not require the data to be normally distributed and is not very sensitive to outliers, so it is more robust. In addition, the Spearman correlation analysis mainly examines the strength of the monotonic relationship between the two variables and can handle nonlinear relationships. Therefore, it is more suitable when dealing with nonlinear variables related to the indoor thermal environment. When constructing the thermal environment model, the Spearman correlation analysis method is first used to quantitatively evaluate the various factors that affect the indoor temperature change, and then the factors with strong correlation are selected as the input variables of the model to predict the indoor temperature change. This can effectively eliminate the interference of variables with less influence, simplify the input of the model, thereby speeding up the training speed of the model and improving the generalization ability of the model.

[0070] From the principle of BP neural network, we know that gradient propagation is an important mechanism for updating weight parameters in the network. However, if the numerical range of training data is too large, the gradient may become very small during back propagation, causing the update speed of weight parameters to become very slow, or even the gradient to disappear. On the contrary, if the numerical range of training data is too small, the gradient may become very large, causing the update speed of weight parameters to become very fast, or even the gradient to explode. Both of these situations will have an adverse effect on the training of the model. By normalizing the data, the numerical range of the input data can be limited, thereby avoiding the occurrence of gradient disappearance or explosion.

[0071] Normalization can also reduce the dependency between different features, which means that the model will not be too sensitive to changes between different features. This makes it easier for the model to capture common features in the data, so that it can be better generalized to new data sets and improve the generalization ability of the model. In addition, normalization can also make the training of the model more stable, because the value range of each feature is limited to a smaller range, and the value of some features will not be too large or too small.

[0072] Therefore, normalization is an important step before training neural network data. It can improve the training speed and accuracy of the model, avoid the disappearance or explosion of gradients, and improve the generalization ability of the model.

[0073] In S2, the BP neural network is optimized using the improved grey wolf optimization algorithm, which includes the following steps:

[0074] S21: Initialize the gray wolf population and set the number of iterations and fitness function;

[0075] S22: According to the set fitness function, the position of the gray wolf population is updated by adopting a dynamic weight update strategy based on the step-length Euclidean distance, and the fitness value of the updated gray wolf population is calculated;

[0076] S23: Determine whether the current iteration meets the set number of iterations. If not, return to step S22. If yes, obtain the optimal individual based on the updated fitness value of the gray wolf population, and use the optimal individual as the initial weight and threshold of the BP neural network for network training, and enter step S24;

[0077] S24: Calculate the output error of the BP neural network and adjust the initial weight and threshold through error back propagation;

[0078] S25: Based on the adjusted initial weights and thresholds, calculate the average error of the BP neural network at the current moment, and determine whether the current iteration meets the preset conditions. If not, return to step S24. If so, output the BP neural network prediction result to complete the optimization of the BP neural network.

[0079] The Gray Wolf Algorithm is widely used to optimize various algorithm models due to its advantages of few set parameters, simple model, and fast optimization speed. However, in the original Gray Wolf Algorithm, since the initial population is randomly generated, the population lacks diversity, and in the iterative process of the Gray Wolf Algorithm, the convergence factor decreases linearly from 2 to 0 as the number of iterations increases. This linear reduction method cannot well balance the global search and local search capabilities of the Gray Wolf Algorithm.

[0080] When solving function optimization problems, the GWO algorithm usually uses randomly generated data as initial information. However, this method of using a pseudo-random number generator to generate an initial population is not easy to produce a diverse population, especially the population generated in a high-dimensional space is not uniform enough. Chaotic motion has the characteristics of randomness, ergodicity, and initial value sensitivity, which can make the algorithm have a faster convergence speed. Adding a chaotic map to generate a chaotic sequence during population initialization can make the initial population as evenly distributed as possible in the solution space, thereby avoiding falling into a local optimum when solving the problem and improving the global search capability. Circle mapping has good uniformity and can generate uniform initial values ​​in [0,1]. Therefore, the present invention uses Circle chaotic mapping to initialize the population.

[0081] In S21, the Circle chaotic map is used to initialize the gray wolf population, and the formula is:

[0082]

[0083] Among them, x m+1 is the chaotic value of the m+1th iteration, x m is the chaotic value of the mth iteration, and mod(·) is the modulo operation.

[0084] From the search mechanism and hierarchy of the GWO algorithm, it can be seen that the final position of the wolf pack, that is, the quality of the solution, is determined by the continuous updating of the positions of the α, β and δ wolves with better fitness values. The existing method simply calculates the average of the position vectors of the leading wolves α, β and δ, and does not fully consider the decision-making ability of the α, β and δ wolves. If the leader α is not the global optimal solution and falls into the local optimal situation, the wolf pack will search in the direction of deviating from the global optimal solution, resulting in local convergence, and it is difficult to ensure the global search ability of the algorithm iteration process. In order to further enhance the search ability of the GWO algorithm and accelerate the optimal convergence speed, the present invention adopts a dynamic weight update strategy based on the step length Euclidean distance.

[0085] The dynamic weight update strategy based on the step length Euclidean distance in S22 is:

[0086]

[0087] Among them, ω 1 ,ω 2 and ω 3 For different weights, X 1 , X 2 and X 3 are different individuals, X(·) is the step length Euclidean distance, and t is the current iteration number.

[0088] By continuously adjusting the weights through the above-mentioned calculation of the modulus of the position vector, a continuous balance between global search capability and local search capability can be achieved, which not only improves the generalization ability of the algorithm, but also accelerates the convergence speed of the wolf pack.

[0089] In S24, the initial weight and threshold are adjusted by error back propagation, and the formula is:

[0090]

[0091] Among them, W j is the weight corresponding to the jth input value, b j is the threshold corresponding to the jth input value, α is the learning rate, and E is the error criterion function that needs to be reduced.

[0092] The convergence factor of GWO has an important influence on the search and hunting of the gray wolf population. The original convergence factor of the gray wolf algorithm is a number that decreases linearly from 2 to 0. However, the convergence process of the gray wolf pack search and hunting algorithm is not linear. Therefore, a simple linear decrease strategy for the convergence factor cannot reflect the search and hunting process of the gray wolf, and cannot completely balance the global and local search capabilities of the algorithm. Therefore, the present invention improves the convergence factor and makes the convergence factor nonlinear.

[0093] The convergence factor r of the improved grey wolf optimization algorithm in S2 is:

[0094]

[0095] Among them, t is the current iteration number and T is the maximum iteration number.

[0096] Convergence factor comparison before and after improvement Figure 2 As shown in the figure, the improved convergence factor has a larger value and a slow decay rate in the early stage of the algorithm iteration, with a larger search range, which is conducive to a large number of global searches; while in the later stage of the iteration, the value is smaller and the decay rate is slowed down again, which is conducive to the wolf pack to concentrate on the target for local search, improving the search accuracy, thereby effectively balancing the global and local search capabilities of the gray wolf algorithm.

[0097] The optimal solution set that the gray wolf optimization algorithm seeks is all the weights and thresholds of the BP neural network. Since the weights and thresholds of the neural network exist in the form of matrices, and the specifications of each matrix are different, only by converting all the weights and thresholds into a vector can unified calculation and processing be performed. This requires encoding the weights and thresholds of the neural network into a vector in a specific format according to the network structure. The length of the encoding should be the sum of the total number of weights and thresholds in the BP neural network. The individuals in the gray wolf population contain all the S 1 weights and S 2The total number S of thresholds is calculated as follows:

[0098] S=S 1 +S 1 =(P×W+W×M)+(W+M)

[0099] Among them, P is the number of input layer nodes, W is the number of hidden layer nodes, and M is the number of output layer nodes.

[0100] After completing the coding of the neural network weight threshold, it is necessary to set the population size of the gray wolf, the number of iterations of the algorithm, and the fitness function. Starting from the initial population, the gray wolf will search, update, and hunt in the solution space. According to the set fitness function, the gray wolf population will be continuously updated, and the fitness value of each individual in the population will be calculated each time. The gray wolf with the best fitness value, that is, the gray wolf with the smallest network training error, will be selected as the three leading wolves α wolf, β wolf, and δ wolf in turn, and the others will form the ω wolf pack.

[0101] In each round of iteration, the three leading wolves, α wolf, β wolf and δ wolf, will guide the ω wolf pack to update its position. After updating the position, the next generation of gray wolf population can be obtained. Then, the fitness value is repeatedly calculated and the gray wolf population is updated. After repeated iterations, the optimal individual, the head wolf α, is finally obtained after the algorithm iteration is completed. This optimal individual is the initial weight and threshold that can minimize the network training error.

[0102] The optimal network weights and thresholds found by the improved gray wolf algorithm are used as the initial weights and thresholds of the BP neural network for network training. Compared with the original BP neural network's randomly set initialization method, the initial weights and thresholds optimized by the improved gray wolf algorithm can significantly reduce the forward training error of the neural network, which is beneficial to the network's back propagation adjustment parameters, speeding up the network training speed while also improving the stability of its overall performance. The trained improved GWO-BP neural network model can be applied to various practical problem studies for more accurate predictions.

[0103] The evaluation indicators in S4 include root mean square error, mean absolute percentage error and complex determination coefficient, and the formula is:

[0104]

[0105] Among them, RMSE is the root mean square error, n is the number of sample data, is the model prediction value of sample i, p i is the actual value of sample i;

[0106]

[0107] Among them, MAPE is the mean absolute percentage error;

[0108]

[0109] Among them, R 2 is the complex determination coefficient, is the average of the actual values ​​of sample i.

[0110] The root mean square error (RMSE) eliminates the impact of dimension by performing a square root operation on the mean square error, making data of different dimensions comparable. The mean absolute percentage error (MAPE) intuitively represents the relative error of the model's prediction in the form of a percentage. 2 ) is a commonly used indicator for evaluating the goodness of fit of a model. It can measure the ability of the model to explain the variation of the dependent variable, that is, how much variation of the dependent variable the model can explain, and its value range is [0,1]. 2 =1, it means that the predicted value of the model is exactly the same as the actual value, which means that the prediction effect of the model is perfect. However, in practical applications, a perfect model does not exist and there is always a certain error. 2 The closer it is to 1, the better the model fit is and the stronger the prediction ability is. If the model error is large, R 2 The value of will be farther and farther away from 1. 2 =0, it means that the model does not explain the variation of the dependent variable at all, that is, the prediction of the model is completely unusable.

[0111] In one embodiment of the present invention, in order to construct an actual thermal environment data-driven model and study and analyze the model performance, this embodiment builds a centralized control experimental platform for a central air-conditioning system. Through this experimental platform, different operating parameters can be set for the central air-conditioning system under different outdoor meteorological parameters to conduct experiments, collect and analyze experimental data, and thus establish an actual thermal environment prediction model based on the data.

[0112] Since the present invention mainly studies the room temperature changes in multiple areas and the operation optimization control of the central air-conditioning system, a central air-conditioning system centralized control experimental platform is built to monitor and control the indoor thermal environment changes and the operation of the air-conditioning equipment in real time. The built central air-conditioning system centralized control experimental platform is mainly composed of air-conditioning cold and heat sources, chilled water distribution system, air-conditioning terminal equipment and automatic control system. The platform is located in the laboratory. The central air-conditioning system controls the heat and humidity environment of the three rooms. The system is as follows Figure 3 shown.

[0113] The main unit of the air conditioner uses an air-cooled heat pump unit as the cold and hot source. The water pump provides power for the operation of the air conditioning water system. During summer cooling, the chilled water is cooled in the air-cooled heat pump unit, and then the water pump sends the chilled water to each air conditioning terminal device. The terminal fan allows the indoor air to fully exchange heat with the chilled water in the terminal heat exchanger, and then sends out cold air to lower the room temperature. At the same time, the cold water flowing through the terminal heat exchanger absorbs the heat of the room and heats up, and finally returns to the air conditioner main unit for further cooling.

[0114] The three experimental rooms where the indoor heat and humidity environment is controlled by the central air-conditioning system have the same size, with a total length of 2.9m×3m, a width of 4.8m, and a height of 3.35m. The room enclosure structure uses 150mm lightweight foam sandwich panels. The return pipe of each terminal equipment branch is installed with an electric valve and an ultrasonic heat meter to monitor the flow of the water system. The setting of the air-conditioning system operating parameters and the collection of experimental data in the experimental platform can be achieved through a centralized control cabinet.

[0115] This embodiment uses the gray wolf optimization algorithm (GWO) to optimize the initial weights and thresholds of the BP neural network, a total of 345, and sets the initial gray wolf population to 50, the maximum number of algorithm iterations to 100, and the fitness function to be the mean square error (MSE) between the network output value and the actual data value. In order to determine the effectiveness of the improved gray wolf optimization algorithm, the GWO-BP neural network and the improved GWO-BP neural network model are run respectively, and the fitness value changes between the GWO and the improved GWO during the gray wolf population update iteration process are compared. Figure 4 shown.

[0116] As the number of iterations increases, the fitness of the gray wolf population is constantly searching and updating in a better direction. The wolf pack of the improved gray wolf optimization algorithm found a better solution than the original gray wolf algorithm when it initially traversed the solution space. After that, it always had a better fitness value during the algorithm iteration process, and finally found the optimal solution in the 69th generation.

[0117] After the gray wolf population searches for the optimal solution, it is substituted into the BP neural network as the initial weight and threshold of the network for training. The indoor thermal environment prediction model established by the BP neural network before optimization, the GWO-BP neural network optimized by the gray wolf algorithm, and the improved GWO-BP neural network optimized by the improved gray wolf algorithm is compared, and the evaluation indicators of different algorithm models are calculated as shown in Table 1.

[0118] Table 1 Performance comparison of room temperature prediction models with different algorithms

[0119]

[0120] It can be seen from Table 1 that the GWO-BP neural network optimized by the Grey Wolf Algorithm is superior to the BP neural network before optimization in all model performance evaluation indicators. The improved GWO-BP neural network model after the Grey Wolf Algorithm is improved has a smaller error and better fit than the algorithm before the improvement, and its performance is significantly better than the traditional BP neural network model.

[0121] The remaining 50 groups of experimental data in the sample data set are used as test samples of the model. The input variable data in the test sample set are input into the established improved GWO-BP neural network model. The prediction performance of the model in practical applications is analyzed by comparing the true value of the output variable in the test sample set and the output prediction value of the model for the test sample input variable. The output variable for comparison is the indoor temperature of the three rooms at the next moment. The results of the comparison between the true value of the room temperature of each room and the predicted value of the improved GWO-BP neural network model are as follows. Figure 5-Figure 7 shown.

[0122] pass Figure 5-Figure 7 It can be seen that the predicted value output by the improved GWO-BP neural network indoor temperature prediction model established by the present invention has a good fit with the true value in the test sample, and the model can well give a relatively accurate predicted value for different input data. In the comparison of 50 groups of test sample data, the MAPE of room temperature prediction in room 1 is 1.33%, room 2 is 1.09%, room 3 is 1.07%, and the total MAPE is 1.16%, of which the maximum relative error is 3.29%, the minimum relative error is 0.01%, and the overall error is small. Therefore, the prediction accuracy of the model is high, and it has excellent prediction performance in practical applications, which can provide a model reference for the indoor thermal environment for the optimization control method of the central air-conditioning system.

[0123] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the invention.

Claims

1. A data-driven method for predicting indoor thermal environment of a central air-conditioning system, characterized in that: The following steps are involved: S1: Screen the variables of thermal environment influencing factors in the building interior and perform data preprocessing on the variables of thermal environment influencing factors; S2: The BP neural network is optimized using the improved grey wolf optimization algorithm to obtain a data-driven indoor thermal environment prediction model for the central air-conditioning system; S3: input the preprocessed thermal environment influencing factor variables into the data-driven central air-conditioning system indoor thermal environment prediction model to predict the indoor thermal environment of the central air-conditioning system; S4: Use the evaluation index to evaluate the performance of the data-driven central air-conditioning system indoor thermal environment prediction model, and complete the data-driven central air-conditioning system indoor thermal environment prediction.

2. A data-driven indoor thermal environment prediction method for a central air-conditioning system according to claim 1, characterized in that: The variables of the factors affecting the indoor thermal environment of the building in S1 include outdoor meteorological data, indoor thermal disturbance data, central air-conditioning operation parameter data and the indoor thermal environment data at the current moment; The outdoor meteorological data includes outdoor temperature and outdoor humidity; The indoor thermal disturbance data includes heat dissipation of personnel, lighting and office equipment; The central air-conditioning operation parameter data includes the water supply temperature of the cold and hot source equipment, the water supply flow rate and air volume of the room terminal equipment; The current indoor thermal environment data includes indoor temperature and indoor humidity.

3. The data-driven indoor thermal environment prediction method for a central air-conditioning system according to claim 1 is characterized in that: In said S1, data preprocessing of the thermal environment influencing factor variables includes correlation analysis and data normalization; The correlation analysis uses the Spearman correlation coefficient ρ s , the formula is: Among them, R(·) is the rank of the variable, x l and l are different variables, the subscript l represents the lth value, and N is the number of variables; The data normalization adopts range normalization or standard deviation normalization.

4. The data-driven indoor thermal environment prediction method for a central air-conditioning system according to claim 1, characterized in that: In S2, the BP neural network is optimized using the improved grey wolf optimization algorithm, which includes the following steps: S21: Initialize the gray wolf population and set the number of iterations and fitness function; S22: According to the set fitness function, the position of the gray wolf population is updated by adopting a dynamic weight update strategy based on the step-length Euclidean distance, and the fitness value of the updated gray wolf population is calculated; S23: Determine whether the current iteration meets the set number of iterations. If not, return to step S22. If yes, obtain the optimal individual based on the updated fitness value of the gray wolf population, and use the optimal individual as the initial weight and threshold of the BP neural network for network training, and enter step S24; S24: Calculate the output error of the BP neural network and adjust the initial weight and threshold through error back propagation; S25: Based on the adjusted initial weights and thresholds, calculate the average error of the BP neural network at the current moment, and determine whether the current iteration meets the preset conditions. If not, return to step S24. If so, output the BP neural network prediction result to complete the optimization of the BP neural network.

5. A data-driven indoor thermal environment prediction method for a central air-conditioning system according to claim 4, characterized in that: In S21, the Circle chaotic map is used to initialize the gray wolf population, and the formula is: Among them, x m+1 is the chaotic value of the m+1th iteration, x m is the chaotic value of the mth iteration, and mod(·) is the modulo operation.

6. A data-driven indoor thermal environment prediction method for a central air-conditioning system according to claim 4, characterized in that: The dynamic weight update strategy based on the step length Euclidean distance in S22 is: Among them, ω1, ω2 and ω3 are different weights, X1, X2 and X3 are different individuals, X(·) is the step length Euclidean distance, and t is the current iteration number.

7. The data-driven indoor thermal environment prediction method for a central air-conditioning system according to claim 4, characterized in that: In S24, the initial weight and threshold are adjusted by error back propagation, and the formula is: Among them, W j is the weight corresponding to the jth input value, b j is the threshold corresponding to the jth input value, α is the learning rate, and E is the error criterion function that needs to be reduced.

8. The data-driven indoor thermal environment prediction method for a central air-conditioning system according to claim 1 is characterized in that: The convergence factor r of the improved grey wolf optimization algorithm in S2 is: Among them, t is the current iteration number and T is the maximum iteration number.

9. The data-driven indoor thermal environment prediction method for a central air-conditioning system according to claim 1, characterized in that: The evaluation indicators in S4 include root mean square error, mean absolute percentage error and complex determination coefficient, and the formula is: Among them, RMSE is the root mean square error, n is the number of sample data, is the model prediction value of sample i, p i is the actual value of sample i; Among them, MAPE is the mean absolute percentage error; Among them, R 2 is the complex determination coefficient, p i is the average of the actual values ​​of sample i.

Images