Air conditioner load prediction method based on characteristic value selection

Abstract

The invention provides an air conditioner load prediction method based on characteristic value selection, and belongs to the technical field of machine learning and building energy saving. According to the air conditioner load prediction method based on feature value selection, feature value selection is carried out through mathematical mutual information and physical redundancy collaboration, the selected feature value is evaluated through a prediction model posterior mode, and the optimal input feature value of the model is accurately determined. According to the air conditioner load prediction method based on characteristic value selection, the dynamic characteristics influencing the real-time load value of the air conditioner can be quantitatively captured, it is ensured that the model has high prediction precision, the precision of the air conditioner load prediction model can be improved by 5.62% on the aspect of correlation coefficients, the average error is reduced by about 16%-22%, and the prediction precision is greatly improved. Meanwhile, the calculation cost is basically not increased, and the method has good practical application value.

Description

Technical Field

[0001] The invention belongs to the technical field of machine learning and building energy conservation, and in particular relates to an air conditioning load prediction method based on characteristic value selection. Background Art

[0002] Accurate prediction of building air conditioning load is crucial to optimizing the operation of air conditioning systems and improving energy management efficiency. In recent years, air conditioning load prediction based on data-driven and machine learning has become a research hotspot. Related research mainly focuses on three aspects: (1) Research on the prediction model itself; (2) Comparative research on optimizing model structure and parameters using optimization algorithms; (3) Research on the impact of eigenvalue selection methods on model performance. Although many scholars have achieved a lot of research results in these areas, there is still a lot of work to be further deepened and improved, especially in the research on eigenvalue selection methods to improve model prediction accuracy. There is still a lot of room for improvement.

[0003] This air conditioning load forecasting method based on black box modeling extracts the features of the actual system historical operation data, self-learns to obtain the mapping relationship between the physical quantity to be predicted (load) and other physical quantities that affect this quantity, completes model training, and can predict the load at future moments. Since these physical quantities that affect the load have dynamic characteristics that change over time, only by accurately selecting the characteristic values ​​of the model input can these dynamic characteristics be quantitatively captured to ensure that the model has a high prediction accuracy. Summary of the invention

[0004] The present invention is made to solve the above-mentioned problem, and aims to provide an air conditioning load prediction method based on characteristic value selection.

[0005] The present invention provides an air conditioning load prediction method based on characteristic value selection, which is used to predict the air conditioning load according to the historical load value L of the air conditioning system. T-m and its corresponding indoor and outdoor characteristics to predict the current load value L of the air conditioning system T , T represents the current moment, m is the time difference, Tm represents the historical moment, and has such characteristics. The feature value selection method based on the coordination of mathematical mutual information and physical redundancy includes the following steps: S10, when the enclosure structure is determined, the historical load value L of the air conditioning system in the set time period is selected T-m The data of indoor and outdoor characteristics corresponding to the sample set are used as the sample set, and the abnormal values ​​and missing values ​​in the sample set are corrected and supplemented respectively, and the data in the sample set are standardized and normalized; S20, the Person correlation coefficient is used to evaluate the historical load value L T-m The indoor and outdoor characteristics and the current load value L TThe correlation is used to analyze the influence on it, so as to eliminate the data that violates the physical principle and obtain the data set; S30, based on Shannon's information entropy theory, calculate the historical load value L in the data set T-m and indoor and outdoor characteristics are respectively related to the current load value L T After sorting the mutual information values, select the historical load values ​​L corresponding to the mutual information values ​​greater than the set value. T-m and / or indoor and outdoor characteristics as characteristic values ​​and form a candidate pool based on them; S40, according to the set time period, the data set is divided into a training set and a test set, and an air-conditioning load prediction model is constructed based on them; S50, considering the mathematical mutual information, 1 to k characteristic values ​​with the largest mutual information values ​​in the candidate pool are selected in turn as input characteristic values ​​for training and prediction of the air-conditioning load prediction model, and k corresponding to the air-conditioning load prediction model with the highest prediction accuracy is obtained, that is, the specific selection of characteristic values ​​based on mathematical mutual information is obtained; S60, considering the physical redundancy, 1 to w characteristic values ​​with the largest mutual information values ​​remaining in the candidate pool and different from the types selected in step S50 are selected in turn as input characteristic values, so as to train and predict the air-conditioning load prediction model with the highest prediction accuracy obtained in step S50, and w corresponding to the air-conditioning load prediction model with the highest prediction accuracy is obtained, that is, the specific selection of characteristic values ​​based on physical redundancy is obtained; S70, according to the optimal air-conditioning load prediction model obtained by training the k+w characteristic values ​​finally determined in steps S50 to S60, the current load value of the air-conditioning system is predicted by inputting historical load values ​​and / or indoor and outdoor characteristics into it.

[0006] The air conditioning load prediction method based on feature value selection provided by the present invention may also have the following features: wherein, in step S10, the indoor and outdoor features include indoor air temperature, indoor air relative humidity, outdoor air temperature and outdoor air relative humidity, the building where the air conditioning system is located is located underground, and in step S20, the positive or negative value of the calculated Person correlation coefficient r represents the historical load value L T-m The indoor and outdoor characteristics and the current load value L T Positively or negatively correlated.

[0007] The air conditioning load prediction method based on characteristic value selection provided by the present invention may also have the following characteristics: wherein step S10 includes the following sub-steps:

[0008] S11, when the enclosure structure is determined, select the historical load value L of the air conditioning system for the set time period T-m and the corresponding indoor and outdoor feature data as the sample set; S12, using the horizontal processing method to correct the continuous outliers of the sample set:

[0009] |Y(d,t)-Y(d,t-1)|>α(t),

[0010] |Y(d,t)-Y(d,t+1)|>β(t),

[0011]

[0012] α(t) and β(t) represent the thresholds of the deviation between the historical load value at time t and the adjacent historical load values, Y(d,t) represents the historical load value at time t on the dth day, Y(d,t-1) and Y(d,t+1) represent the historical load values ​​at time t-1 and time t+1 on the dth day, respectively.

[0013] S13, using the vertical processing method, based on the similar characteristics of historical load values ​​at the same time on similar days, the abnormal values ​​are replaced with data from similar dates at the same time:

[0014] |Y(d,t)-a(t)|>r(t),

[0015]

[0016] a(t) represents the average value of the historical load value of similar days at time t, and r(t) represents the threshold between the historical load value at time t and the average value;

[0017] S14, for single missing value filling, the average interpolation method is used to fill the gap by calculating the average of the two data before and after the missing value. For two or more consecutive missing values, the cubic spline interpolation method is used: It is known that the function f(x) has n+1 different nodes on the interval [a, b], a=x0 <x1<…<x i <… <x n =The function value at b is y i =f(x i ), the constructor s(x) satisfies [x i ,x i +1] is 3, and s(x) and the first and second order derivatives are continuous on [a,b]. The calculation formula is c, d, e, and f are polynomial coefficients;

[0018] S15, normalize the sample set, scale the overall data to [0,1], reduce the order of magnitude dimension gap between different data types, and eliminate abnormal results: Y Normalization Represents the normalized data, Y i represents the i-th data, Y max , Y min Respectively represent the maximum and minimum values ​​of all data;

[0019] S16, standardize the sample set and transform the data into a normal distribution with a mean of 0 and a standard deviation of 1, so as to better map the relationship between the physical quantity to be predicted and the input features: Y Standardization represents the data after standardization, μ and σ represent the mean and variance of the data respectively.

[0020] In the air conditioning load prediction method based on feature value selection provided by the present invention, the following features can also be provided: wherein step S20 includes the following sub-steps: S21, using the Person correlation coefficient r to evaluate the historical load value and the indoor and outdoor features respectively and the current load value L T Relevance: N represents the number of feature samples, α j and β j They respectively represent the characteristic variables between the j-th historical load value and the influencing factors, the value range of r is [-1,1], r in the range of [-1,0] represents that the two variables are linearly negatively correlated, r in the range of [0,1] represents that the two variables are linearly positively correlated, |r| tends to 1 represents that the correlation between the two variables is high, and |r| tends to 0 represents that the correlation is low; S22, respectively calculate the correlation coefficient r between several historical load values ​​and / or indoor and outdoor characteristics at the past historical moments and the current load value, and eliminate the historical load values ​​and / or indoor and outdoor characteristics in the sample set corresponding to the calculated correlation coefficient r that violates the physical principle.

[0021] In the air conditioning load prediction method based on feature value selection provided by the present invention, it can also have the following characteristics: wherein step S30 includes the following sub-steps: S31, based on Shannon's information entropy theory, using the mutual information calculation formula to calculate the mutual information values ​​of the historical load values ​​and the indoor and outdoor features in the data set about the current load value: G represents the number of information events, u and v represent the uth and vth information events respectively, P(X u ,Y v ) indicates the Xth u ,Y v The joint probability of the occurrence of information events, P(X u ) and P(Y v ) represent the Xth u and Y v The probability of an information event occurring is calculated, and the mutual information value is used to quantitatively describe the correlation between the two event sets. The larger the mutual information value, the greater the degree of mutual dependence of the two variables and the stronger the correlation. S32, the mutual information values ​​calculated in step S31 are combined and arranged in descending order, and the historical load values ​​and / or indoor and outdoor features corresponding to all mutual information values ​​not less than the set value are selected as feature values ​​and arranged in descending order to form a candidate pool.

[0022] The air conditioning load prediction method based on characteristic value selection provided by the present invention may also have the following characteristics: wherein, in step S32, the set value is 0.185.

[0023] The air conditioning load prediction method based on feature value selection provided by the present invention may also have the following feature: wherein, in step S40, the air conditioning load prediction model is constructed by a long short-term memory network or a BP neural network.

[0024] The air conditioning load prediction method based on characteristic value selection provided by the present invention may also have the following characteristics: wherein, in steps S50 to S60, the prediction accuracy is calculated in the following manner: Correlation coefficient Mean absolute error Root mean square error Mean absolute percentage error For the predicted value The covariance with the actual value γ, and Var[γ] are the predicted values The variance of the actual value γ, n is the prediction duration, and γ z are the predicted value and actual value at the zth moment respectively. The smaller the value of ξ, the lower the prediction accuracy. The larger the values ​​of MAE, RMSE and MAPE, the lower the prediction accuracy.

[0025] The air conditioning load prediction method based on feature value selection provided by the present invention may also have the following characteristics: wherein, in steps S50 to S60, when adding feature values ​​of the same type as input feature values ​​to train and predict the air conditioning load prediction model and the prediction accuracy decreases, the increase is stopped, otherwise the increase is continued until the prediction accuracy begins to decrease; when adding feature values ​​of the same type as input feature values ​​to train and predict the air conditioning load prediction model until the prediction accuracy begins to decrease, feature values ​​of different types are started to be added as input feature values ​​to train and predict the air conditioning load prediction model; if the prediction accuracy improves, the increase is continued, otherwise the increase is stopped.

[0026] Functions and Effects of the Invention

[0027] According to the air conditioning load prediction method based on eigenvalue selection involved in the present invention, the eigenvalues ​​of the model input are accurately selected because of the eigenvalue selection method based on the coordination of mathematical mutual information and physical redundancy.

[0028] Therefore, the air conditioning load prediction method based on characteristic value selection of the present invention can quantitatively capture the dynamic characteristics that affect the real-time load value of the air conditioner, ensure that the model has a high prediction accuracy, and can improve the accuracy of the air conditioning load prediction model by 5.62% in the correlation coefficient and reduce the average error by about 16% to 22%. At the same time, it basically does not increase the calculation cost, and has good practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flowchart of an air conditioning load prediction method based on feature value selection according to an embodiment of the present invention;

[0030] Figure 2 is the hourly historical air conditioning load value in the sample set after the data preprocessing in step S10 in the embodiment of the present invention;

[0031] Figure 3 is the historical load value L calculated in step S22 of the embodiment of the present invention. T-m , indoor air temperature, indoor air relative humidity, outdoor air temperature and outdoor air relative humidity in the past 24 hours (m = 0 ~ 24) and the current load value L T The Person correlation coefficient r change curve between them;

[0032] Figure 4 is a candidate pool of feature values ​​obtained by merging the mutual information values ​​in the embodiment of the present invention and arranging them in descending order;

[0033] Figure 5 is a sample diagram of a training set of a feature value in step S50 of an embodiment of the present invention;

[0034] Figure 6 is a diagram of training set samples of two feature values ​​in step S50 of an embodiment of the present invention;

[0035] Figure 7 is the influence of the number of characteristic values ​​on the prediction accuracy in step S50 of the embodiment of the present invention;

[0036] Figure 8 is the air conditioning load prediction result under 6 scenarios in step S50 of the embodiment of the present invention;

[0037] Fig. 9 is a box plot of relative errors of prediction results under 6 scenarios in step S50 of an embodiment of the present invention;

[0038] Fig.10 It is the accuracy comparison and performance evaluation index of the air conditioning load prediction model when the characteristic value type is the historical load value and the outdoor air temperature in step S60 of the embodiment of the present invention;

[0039] Fig.11 It is the accuracy comparison and performance evaluation index of the air conditioning load prediction model when the characteristic value types in step S60 of the embodiment of the present invention are historical load values, outdoor air temperature and indoor air temperature;

[0040] Fig.12 It is the accuracy comparison and performance evaluation index of the air conditioning load prediction model when the characteristic value types in step S60 of the embodiment of the present invention are historical load values, outdoor air temperature, indoor air temperature and outdoor air relative humidity. DETAILED DESCRIPTION

[0041] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following embodiments and drawings specifically illustrate an air conditioning load prediction method based on characteristic value selection and its application of the present invention.

[0042] <Example>

[0043] Figure 1 It is a flowchart of an air conditioning load prediction method based on feature value selection according to an embodiment of the present invention.

[0044] like Figure 1 As shown, this embodiment provides an air conditioning load prediction method based on eigenvalue selection, and an eigenvalue selection method based on the coordination of mathematical mutual information and physical redundancy, which is used to predict the air conditioning load according to the historical load value L of the air conditioning system. T-m and its corresponding indoor and outdoor characteristics to predict the current load value L of the air conditioning system T , T represents the current time, m represents the time difference, Tm represents the historical time, including the following steps S10 to S60:

[0045] S10, when the enclosure structure is determined, obtain a sample set and pre-process the data therein, including the following sub-steps S11 to S16:

[0046] S11, select the actual operation data of an underground building air-conditioning system determined by the enclosure structure in the hot-summer and warm-winter area from 0:00 on May 1, 2020 to 23:00 on August 31, 2020 as the sample set. The air-conditioning system has a total of 2 chillers.

[0047] The sample set contains the historical load values ​​L of the two chillers in the air-conditioning system during the set time period. T-m And its corresponding indoor and outdoor feature data.

[0048] Among them, the factors affecting the building air conditioning load include the envelope structure and indoor and outdoor characteristics. Once the envelope structure is determined, its dynamic impact on the air conditioning load is small; while indoor and outdoor characteristics include indoor air temperature and humidity, indoor personnel, equipment and lighting, outdoor air temperature and humidity, outdoor wind speed and solar radiation illumination, which often have a greater impact on the dynamic changes of the air conditioning load.

[0049] The building in this embodiment is located underground, so the solar radiation and outdoor wind speed parameters have little effect on the building air conditioning load. Therefore, the indoor and outdoor features selected in this embodiment are the indoor air temperature at 4 collection points in the underground building, the indoor air relative humidity at 4 collection points in the underground building, the outdoor air temperature at the entrance of the underground building, and the outdoor air relative humidity at the entrance of the underground building. In the above sample set, the data sampling interval is 2 minutes, and a total of 88,560 sets of data.

[0050] When the air conditioning system is running, under natural conditions, due to the accuracy of sensors, systematic errors in measuring instruments and data collected by personnel, sudden instability of operating conditions, etc., the sample set inevitably often contains abnormal values. If it is directly used to train the prediction model, it will inevitably affect the prediction accuracy of the model and lead to inaccurate results. Therefore, the data in the sample set should be preprocessed through the following steps S12 to S14.

[0051] S12, using the horizontal processing method, corrects the continuous outliers in the sample set:

[0052] |Y(d,t)-Y(d,t-1)|>α(t), |Y(d,t)-Y(d,t+1)|>β(t),

[0053] Among them, α(t) and β(t) represent the threshold values ​​of the deviation between the historical load value at time t and the adjacent historical load values, Y(d,t) represents the historical load value at time t on the d day, Y(d,t-1) and Y(d,t+1) represent the historical load values ​​at time t-1 and time t+1 on the d day, respectively.

[0054] S13, using the vertical processing method, based on the similar characteristics of historical load values ​​at the same time on similar days, the abnormal values ​​are replaced by data from similar dates at the same time: |Y(d,t)-a(t)|>r(t),

[0055]

[0056] Where a(t) represents the average value of the historical load values ​​of similar days at time t, and r(t) represents the threshold between the historical load value at time t and the average value.

[0057] S14, fill in the missing values ​​in the sample set:

[0058] (1) To fill in a single missing value, the average interpolation method is used to fill in the gap by calculating the average of the two data before and after the missing value.

[0059] (2) To fill in two or more consecutive missing values, use the cubic spline interpolation method:

[0060] It is known that the function f(x) has n+1 distinct nodes on the interval [a,b], a=x0 <x1<…<x i <… <x n =The function value at b is y i =f(x i ), the constructor s(x) satisfies [x i ,x i +1] is 3, and s(x) and the first and second order derivatives are continuous on [a,b]. The calculation formula is Among them, c, d, e, and f are polynomial coefficients.

[0061] S15, normalize the sample set, scale the overall data to [0,1], reduce the order of magnitude dimension gap between different data types, and eliminate abnormal results:

[0062] Among them, Y Normalization Represents the normalized data, Y i represents the i-th data, Y max , Y min Respectively represent the maximum and minimum values ​​of all data;

[0063] S16, standardize the sample set and transform the data into a normal distribution with a mean of 0 and a standard deviation of 1, so as to better map the relationship between the physical quantity to be predicted and the input features:

[0064] Among them, Y Standardization represents the data after standardization, μ and σ represent the mean and variance of the data respectively.

[0065] Figure 2 It is the hourly historical air conditioning load value in the sample set after the data preprocessing in step S10 in the embodiment of the present invention.

[0066] like Figure 2 As shown, after the relevant data processing of steps S11 to S16, the data in the sample set has continuity and can be used for black box prediction of air conditioning load.

[0067] S20, removing data that violates physical principles from the sample set to obtain a data set, including the following sub-steps S21-S22:

[0068] S21, the Person correlation coefficient r is used to evaluate the historical load value and the indoor and outdoor characteristics and the current load value L T Relevance:

[0069] Among them, N represents the number of feature samples, α j and β j They represent the characteristic variables between the j-th historical load value and the influencing factor respectively. The value range of r is [-1,1]. r in the range of [-1,0] indicates that the two variables are linearly negatively correlated, r in the range of [0,1] indicates that the two variables are linearly positively correlated, |r| tends to 1, which means that the correlation between the two variables is high, and |r| tends to 0, which means that the correlation is low.

[0070] S22, due to thermal inertia and time delay, the historical load value L T-m and indoor and outdoor characteristics (indoor air temperature, indoor air relative humidity, outdoor air temperature and outdoor air relative humidity) at different historical moments to the current air conditioning load (current load value L T ) have different influences, and SPSS Statistics 21 software is used to calculate five characteristic physical quantities (historical load value L T-m , indoor air temperature, indoor air relative humidity, outdoor air temperature and outdoor air relative humidity) in the past 24 hours (m = 0 ~ 24) and the current load value L T The Person correlation coefficient r between them is as follows: Figure 3 As shown. Among them, Figure 3 is the historical load value L calculated in step S22 of the embodiment of the present invention. T-m , indoor air temperature, indoor air relative humidity, outdoor air temperature and outdoor air relative humidity in the past 24 hours (m = 0 ~ 24) and the current load value L T The Person correlation coefficient r change curve between them.

[0071] According to the physical relationship between indoor air conditioning load and indoor air temperature, indoor air relative humidity, outdoor air temperature and outdoor air relative humidity, the correlation rule judgment condition is given as follows: in the summer air conditioning operation season, the indoor air temperature, outdoor air temperature and current load value L T There is a positive correlation, r is a positive value; the relative humidity of indoor air, the relative humidity of outdoor air and the current load value L T There is a negative correlation, and r is a negative value.

[0072] The r value of outdoor air temperature is negative from T-9 to T-17, the r value of outdoor air relative humidity is positive from T-4 to T-20, and the r value of indoor air relative humidity is positive from T to T-24, all of which violate the regularity judgment condition. Therefore, these characteristic physical quantities in these time periods in the original data set have no significant effect on the air conditioning load at the current time (current load value L T ) is a phenomenon that violates the principles of physics, so these original data are eliminated.

[0073] Among them, the specific data eliminated are: outdoor air temperature from historical moments T-9 to T-17, outdoor air relative humidity from historical moments T-4 to T-20, and indoor air relative humidity from historical moments T to T-24.

[0074] After removing the above data from the sample set, we get the data set.

[0075] S30, select the data set that can be used to calculate the current load value L T The historical load value L that has a significant impact T-m and / or indoor and outdoor features as feature values ​​and forming a candidate pool, including the following sub-steps S31-S32:

[0076] S31, based on Shannon’s information entropy theory, uses the mutual information calculation formula to calculate the mutual information values ​​of the historical load values ​​and indoor and outdoor features in the data set about the current load value:

[0077]

[0078] Among them, G represents the number of information events, u and v represent the uth and vth information events respectively, and P(X u ,Y v ) indicates the Xth u ,Y v The joint probability of the occurrence of information events, P(X u ) and P(Y v ) represent the Xth u and Y v The mutual information value is used to quantitatively describe the correlation between two event sets. The larger the mutual information value, the greater the degree of mutual dependence between the two variables and the stronger the correlation.

[0079] Quantitative calculation of the historical load values ​​L of the four characteristic physical quantities (m = 1 to 24) in the past 24 hours T-m , indoor air temperature from time T to T-24, outdoor air temperature from time T to T-8 and from time T-18 to T-24, outdoor air relative humidity from time T to T-3 and from time T-21 to T-24) and the current load value L T The mutual information value of .

[0080] Using Python programming language, the continuous function under the entropy_estimators self-editing library is called to calculate the mutual information value. The mutual information value results are shown in Table 1 below.

[0081] Table 1 (Mutual information values ​​of historical load values, indoor air temperature, outdoor air temperature and outdoor air relative humidity in the past 24 hours about the current load value in the data set)

[0082]

[0083] S32, since the mutual information value reflects the degree of mutual dependence between two variables, the calculation conditions of all mutual information values ​​in Table 1 in step S31 are the same, and they are comparable and consistent with each other. Therefore, all mutual information values ​​in Table 1 reflect the correlation between the historical load values ​​and indoor and outdoor characteristics at different times in the past 24 hours and the current load values. The larger the mutual information value, the greater the influence on the current load value, and it is used as the input characteristic value of the air-conditioning load prediction model.

[0084] Figure 4 It is a candidate pool of feature values ​​obtained by merging the mutual information values ​​in the embodiment of the present invention and arranging them in descending order.

[0085] like Figure 4 As shown, all mutual information values ​​in Table 1 in step S31 are merged and arranged in descending order, and all historical load values ​​and indoor and outdoor characteristics (indoor air temperature, outdoor air temperature and outdoor air relative humidity) with mutual information values ​​greater than or equal to 0.185 are taken as feature values, thereby forming a candidate pool of input feature values ​​for the air-conditioning load prediction model of the building.

[0086] like Figure 4 As shown in the figure, the horizontal axis represents all the eigenvalues ​​from left to right, and their influence on the air conditioning load decreases from left to right. Therefore, all the eigenvalues ​​in the candidate pool have reasons to be selected as the input eigenvalues ​​of the air conditioning load prediction model. However, if Figure 4 If all the eigenvalues ​​of are used to train the model, it will not only greatly increase the parameter collection cost in the actual system and the calculation cost of the prediction model, but also due to the overfitting characteristics of the machine learning model, it is likely to lead to low final prediction accuracy of the model.

[0087] Therefore, how to select the optimal characteristic value is the key to improving the accuracy of the air conditioning load prediction model. Therefore, this embodiment performs the optimal selection of the characteristic value through the following steps S40 to S60.

[0088] S40, with hours as the time granularity, divide the original data set obtained in step S20 into a training set and a test set, and use the data from 0:00 on May 1 to 23:00 on August 29 as the training set; and use the data from 0:00 on August 30 to 23:00 on August 31 as the test set.

[0089] The long short-term memory network (LSTM) is used to build an air conditioning load forecasting model (LSTM model). According to multiple simulation experiments, the model-related parameter settings are shown in Table 2. The MATLAB R2019a programming language is used to call the Deep Learning related library in Simulink to write the LSTM model algorithm.

[0090] Table 2 (Prediction model related parameter settings)

[0091]

[0092] The model prediction evaluation indicators used by the air conditioning load prediction model in this embodiment are: correlation coefficient ξ, mean absolute error MAE, root mean square error RMSE, mean absolute percentage error MAPE, and the calculation formula is as follows:

[0093] Correlation coefficient

[0094] Mean absolute error

[0095] Root mean square error

[0096] Mean absolute percentage error

[0097] in, For the predicted value The covariance with the actual value γ, and Var[γ] are the predicted values The variance of the actual value γ, n is the prediction duration, and γ z are the predicted value and actual value at the zth moment respectively. The smaller the value of ξ, the lower the prediction accuracy. The larger the values ​​of MAE, RMSE and MAPE, the lower the prediction accuracy.

[0098] S50, select the eigenvalue based on the mathematical mutual information, select 1 to k eigenvalues ​​with the largest mutual information values ​​in the candidate pool as input eigenvalues ​​to train and predict the air conditioning load prediction model, and obtain the k corresponding to the air conditioning load prediction model with the highest prediction accuracy, that is, obtain the specific selection of the eigenvalue based on the mathematical mutual information. The specific method of this step is as follows:

[0099] The air conditioning load prediction model constructed in step S40 is used to Figure 4 Various eigenvalue selection methods with mutual information values ​​ranking top 1, top 2, top 3, top 4, top 5 and top 6 in the candidate pool shown in the figure are used for simulation experiments, and a total of 6 experimental scenarios are set, as shown in Table 3 below.

[0100] Table 3 (6 experimental scene settings)

[0101]

[0102] Taking scenario 1 as an example, the first LSTM network model is constructed, and the relevant parameter settings are shown in Table 2 in step S40. Figure 3 The first quantity on the left in the candidate pool, i.e. the historical air conditioning load L at time T-1 T-1 As the input feature value (abbreviated as "one feature value"), a data set is constructed to train and predict the model.

[0103] Figure 5 is a diagram of a training set sample of a feature value in step S50 of an embodiment of the present invention.

[0104] like Figure 5 As shown, the data of the training set is input into the model with a matrix dimension of 2904×1 for training. After the iterative calculation converges, the data in the test set is tested with a matrix dimension of 48×1, and the corresponding accuracy index is calculated.

[0105] Similarly, in scenario 2, the second LSTM network model is constructed, and the input feature value is Figure 3 The first two quantities on the left in the candidate pool.

[0106] Figure 6 is a diagram of training set samples of two feature values ​​in step S50 of an embodiment of the present invention.

[0107] like Figure 6 As shown, the method of scenario 2 is similar, but it is necessary to check the missing values ​​of the data matrix in the training set and the test set respectively ( Figure 6 The historical moments in yellow are supplemented.

[0108] According to the feature values ​​selected in Table 3, the LSTM model is used to conduct experiments to predict the air conditioning load of the underground building. Figure 4 The first six characteristic physical quantities from left to right (L T-1 , L T-2 , L T-3 , L T-4 , L T-5 , L T-6 ).

[0109] Figure 7is the influence of the number of characteristic values ​​on the prediction accuracy in step S50 of the embodiment of the present invention; Figure 8 is the air conditioning load prediction result under 6 scenarios in step S50 of the embodiment of the present invention; Fig. 9 It is a box plot of relative errors of prediction results in six scenarios in step S50 of an embodiment of the present invention.

[0110] like Figure 7 to Figure 9 As shown in the figure, before the number of eigenvalues ​​is increased to four, the prediction evaluation indicators of various errors are continuously reduced, and the prediction accuracy is gradually improved; but after the eigenvalue exceeds four, if the number of eigenvalues ​​is continued to increase, the evaluation indicators of various errors will increase with the increase of eigenvalues, and the prediction accuracy will begin to decrease.

[0111] This is because too few eigenvalues ​​(k) will cause the model to be unable to fully learn the input features, resulting in underfitting during the training process, which will lead to low prediction accuracy of the model; while too many eigenvalues ​​will increase the maximum and minimum redundancy of the parameters, and there will be competition between multiple groups of input parameters of the same category. After training, the historical load input model with a small mutual information value will affect the historical load with a large mutual information value, causing the prediction accuracy of the model with the original input quantitative eigenvalue to decrease.

[0112] Taking the LSTM model as an example, when there are too many input feature values, the model becomes more complex, the training time increases, and the cost of collecting data increases. The evaluation indicators and calculation time τ of the load forecasting model under 6 scenarios are shown in Table 4 below.

[0113] Table 4 (Evaluation index and calculation time τ of load forecasting model under 6 scenarios)

[0114]

[0115] like Figure 7 , Figure 8 , Fig. 9 As shown in Table 4, for the air conditioning load prediction model, the number of eigenvalues ​​selected according to the mutual information ranking is not the more the better, but there is an optimal number of eigenvalues ​​selected k. If this number is exceeded, the model prediction performance may be disadvantageous. This embodiment found through simulation experiments that the optimal number of eigenvalues ​​is four (i.e. k = 4, and the eigenvalues ​​with the top four mutual information values ​​are L T-1 , L T-2 , L T-3 , L T-4 ), and the fourth eigenvalue L T-4 The mutual information value between the predicted load and the model is 0.363. At the same time, the optimal number of eigenvalues ​​k = 4 is selected, and the prediction time of the model does not increase, and the calculation cost does not increase.

[0116] from Figure 4 Judging from the mutual information value of the candidate pool, if we only look at the mathematical mutual information value, the ones at the top are the historical load values ​​(black columns).

[0117] As can be seen from step S50, the best prediction accuracy is achieved by selecting the first four eigenvalues. If the number is increased to the first five (when k=5), the prediction accuracy decreases. Figure 3 The indoor air temperature, outdoor air temperature or indoor air relative humidity (red, yellow or green columns) ranked lower in the list are generally considered to be unsuitable to be selected as the input feature values ​​of the model due to their small mutual information values.

[0118] However, the current load value of the air conditioning in an actual underground building is related to the personnel load, indoor lighting and equipment heat generation, and underground soil heat storage. Therefore, the current load value of the air conditioning will also be affected by indoor and outdoor air parameters. Therefore, from the perspective of the load generation mechanism, indoor and outdoor temperature and humidity will also affect the building air conditioning load.

[0119] Therefore, in Figure 4 In the candidate pool, it is also necessary to select other types of physical quantities (indoor air temperature, outdoor air temperature or outdoor air relative humidity) that may affect the current load value of the air conditioner and are not historical load values ​​as feature values, while taking into account the numerical sorting of their mutual information, so the following step S60 is performed:

[0120] S60, under the coordination of physical redundancy, select the 1 to w eigenvalues ​​with the largest mutual information values ​​remaining in the candidate pool and different from the type selected in step S50 as input eigenvalues, so as to train and predict the air-conditioning load prediction model with the highest prediction accuracy obtained in step S50 again, and obtain w corresponding to the air-conditioning load prediction model with the highest prediction accuracy (better than step S50), that is, obtain the specific selection of eigenvalues ​​based on the coordination of physical redundancy. The specific method of this step is as follows:

[0121] As shown in Table 5, based on the number of the four best eigenvalues ​​selected in step S50, the outdoor air temperature C with the highest mutual information value among the other categories of characteristic physical quantities (outdoor air temperature) other than the historical load value is selected. T,o (o represents outdoor, T represents the current time, and C represents temperature) as the fifth eigenvalue ( Figure 4 The selected eigenvalues ​​are still evaluated in a posterior manner using the LSTM model until the number of eigenvalues ​​for the optimal outdoor air temperature is obtained.

[0122] Continue to add other categories of characteristic physical quantities (indoor air temperature) as characteristic values ​​( Figure 4Similarly, after selecting the best eigenvalue (indoor air temperature), add the outdoor air relative humidity as the eigenvalue ( Figure 4 The first green column that appears in the prediction) is used until the prediction accuracy of the air-conditioning load prediction model reaches the highest, and the number of eigenvalues ​​with the largest mutual information value different from the type selected in step S50 that should be selected in this step can be obtained.

[0123] Table 5 (Types and quantities of different categories of feature values ​​added in step S60)

[0124]

[0125] In the above Table 5, L represents the load of the air conditioner, T represents the current time, Tm represents the historical time (m=1-24), C represents the temperature, o represents the outdoor, q represents the indoor, and RH represents the relative humidity of the air.

[0126] In this step, the LSTM model posterior method is still used to conduct simulation experiments on various scenarios in Table 5.

[0127] Fig.10 It is the accuracy comparison and performance evaluation index of the air conditioning load prediction model when the characteristic value type is the historical load value and the outdoor air temperature in step S60 of the embodiment of the present invention; Fig.11 It is the accuracy comparison and performance evaluation index of the air conditioning load prediction model when the characteristic value types in step S60 of the embodiment of the present invention are historical load values, outdoor air temperature and indoor air temperature; Fig.12 It is the accuracy comparison and performance evaluation index of the air conditioning load prediction model when the characteristic value types in step S60 of the embodiment of the present invention are historical load values, outdoor air temperature, indoor air temperature and outdoor air relative humidity.

[0128] like Figure 10 to Figure 12 The accuracy comparison and performance evaluation index of the air conditioning load prediction model under different eigenvalue types are shown as follows:

[0129] (1) When the eigenvalues ​​are selected as the top five historical load values ​​sorted by mutual information value, the prediction accuracy of the model begins to decrease. However, if the physical redundancy is considered while taking into account the mathematical mutual information sorting, C, which is only ranked 13th in the total mutual information sorting but has an impact on the current load value and is ranked first in the characteristic physical quantities such as outdoor temperature, will be T,o If we select C as the fifth eigenvalue, the error indexes of the model prediction will be reduced, and the model accuracy will be further improved. However, if we select C as the fifth eigenvalue, the error indexes of the model prediction will be reduced, and the model accuracy will be further improved. T-1,o If it is selected as the sixth eigenvalue, the model prediction accuracy decreases.

[0130] (2) C, which ranks only 16th in the total mutual information value ranking but ranks first in other categories of characteristic physical quantities such as indoor temperature, is ranked first in the total mutual information value ranking. T-18,q As the sixth eigenvalue, it can also further improve the final prediction accuracy of the model. However, if this type of physical quantity is added, it will not be conducive to the final prediction accuracy of the model.

[0131] (3) However, if the fourth type of characteristic physical quantity, outdoor relative humidity, is ranked first, RH T-24,o As the seventh eigenvalue, it cannot bring further improvement in the prediction accuracy of the model. This is because although RH T-24,o Characteristic physical quantities belonging to other categories different from the previous ones, but which are too far behind in the ranking of the total mutual information value, have too little correlation with the predicted load.

[0132] It can be seen that while taking into account the mathematical quantitative indicators, selecting parameters with relatively low total ranking but high mutual information values ​​among other categories of physical quantities as eigenvalues ​​can help improve the final prediction accuracy of the model. This can be explained from the physical redundancy influence of the model input eigenvalues. Selecting outdoor temperature with a small mutual information value but different from the historical load physical quantity type as the eigenvalue can avoid the maximum and minimum redundancy between the same type of physical quantities, and the model can extract the features of other categories of physical quantities to more fully explore the mapping relationship between historical data and predicted quantities, thereby helping to improve the prediction accuracy of the model. However, when the second outdoor temperature is added, the prediction accuracy of the model shows a downward trend again, indicating that adding too many outdoor temperatures of the same type will increase the competition relationship of the same type, and the model cannot effectively learn related features, resulting in reduced prediction accuracy.

[0133] However, while considering the physical redundancy, it is also necessary to consider the total ranking of the characteristic physical quantity in the correlation quantification index. In this embodiment, the outdoor air relative humidity RH T-24,o The lower the total ranking is, the lower the correlation is with the current load value to be predicted. Therefore, it cannot provide more valuable data features for the prediction model. If it is selected as a feature value, the prediction accuracy will be reduced.

[0134] like Figure 10 to Figure 12 The accuracy comparison and performance evaluation index of the air conditioning load prediction model under different eigenvalue types are shown in Figure 4When the mutual information rankings in the candidate pool are the six characteristic values ​​of the historical loads at T-1, T-2, T-3, and T-4, the outdoor air temperature at T, and the indoor air temperature at T-18, which are ranked 1st to 4th, 13th, and 16th respectively (i.e., k=4 in step S50 and w=2 in step S60), the accuracy evaluation indicators of the LSTM air-conditioning load prediction model are: correlation coefficient 0.94, average absolute error 15.75kW, root mean square error 22.03kW, and average absolute percentage error 3.05%. Compared with the model accuracy evaluation indicators when only the historical loads at T-1, T-2, T-3, and T-4 at the top four in the total ranking are selected as characteristic values, the correlation coefficient is increased by 5.62%, the average absolute error is reduced by 16.13%, the root mean square error is reduced by 22.76%, and the average absolute percentage error is reduced by 16.44%. The slightly extended calculation time hardly increases the calculation cost, which is negligible compared with the improvement of the model prediction accuracy.

[0135] S70, k+w eigenvalues ​​(specifically, in this embodiment, Figure 4 The optimal air-conditioning load prediction model is obtained by training the six characteristic values ​​of historical loads at time T-1, T-2, T-3, and T-4, outdoor air temperature at time T, and indoor air temperature at time T-18, whose mutual information rankings in the candidate pool are 1 to 4, 13, and 16 respectively. The historical load values ​​and / or indoor and outdoor characteristics are input into the model to predict the current load value of the air-conditioning system.

[0136] Functions and Effects of the Embodiments

[0137] According to the air conditioning load prediction method based on eigenvalue selection involved in this embodiment, the optimal input eigenvalue of the model is accurately determined because of the eigenvalue selection method based on the coordination of mathematical mutual information and physical redundancy, and the selected eigenvalue is evaluated by the LSTM model posteriori.

[0138] Therefore, the air conditioning load prediction method based on characteristic value selection of the present invention can quantitatively capture the dynamic characteristics that affect the real-time load value of the air conditioner, ensure that the model has a high prediction accuracy, and can improve the accuracy of the air conditioning load prediction model by 5.62% in the correlation coefficient and reduce the average error by about 16% to 22%. At the same time, it basically does not increase the calculation cost, and has good practical application value.

[0139] Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An air conditioning load prediction method based on characteristic value selection is used to predict the air conditioning load according to the historical load value L of the air conditioning system. T-m and its corresponding indoor and outdoor characteristics to predict the current load value L of the air conditioning system T , T represents the current moment, m is the time difference, Tm represents the historical moment, It is characterized in that The eigenvalue selection method based on the coordination of mathematical mutual information and physical redundancy includes the following steps: S10, when the enclosure structure is determined, select the historical load value L of the air conditioning system during the set time period T-m and the corresponding data of the indoor and outdoor features as a sample set, and after respectively correcting and supplementing the abnormal values ​​and missing values ​​in the sample set, standardizing and normalizing the data in the sample set; S20, using the Person correlation coefficient to evaluate the historical load value L T-m The indoor and outdoor characteristics and the current load value L T The correlation is used to analyze the influence on it, so as to eliminate the data that violates the physical principle and obtain the data set; S30, based on Shannon's information entropy theory, calculate the historical load value L in the data set T-m and the indoor and outdoor characteristics are respectively related to the current load value L T After sorting the mutual information values, select the historical load values ​​L corresponding to the mutual information values ​​greater than the set value. T-m and / or the indoor and outdoor features are used as feature values ​​to form a candidate pool; S40, dividing the data set into a training set and a test set according to the set time period, and constructing an air conditioning load prediction model based on the training set and the test set; S50, considering mathematical mutual information, sequentially selecting 1 to k eigenvalues ​​with the largest mutual information values ​​in the candidate pool as input eigenvalues ​​to train and predict the air conditioning load prediction model, and obtaining k corresponding to the air conditioning load prediction model with the highest prediction accuracy, that is, obtaining a specific selection of the eigenvalue based on mathematical mutual information; S60, considering physical redundancy, sequentially selecting the 1 to w eigenvalues ​​with the largest mutual information values ​​remaining in the candidate pool and different from the types selected in step S50 as input eigenvalues, so as to train and predict the air-conditioning load prediction model with the highest prediction accuracy obtained in step S50, and obtaining w corresponding to the air-conditioning load prediction model with the highest prediction accuracy, that is, obtaining the specific selection of the eigenvalue based on physical redundancy; S70, predicting the current load value of the air-conditioning system by inputting historical load values ​​and / or indoor and outdoor characteristics into the optimal air-conditioning load prediction model obtained by training the k+w characteristic values ​​finally determined in steps S50 to S60.

2. The air conditioning load prediction method based on characteristic value selection according to claim 1 is characterized in that: in, In step S10, the indoor and outdoor characteristics include indoor air temperature, indoor air relative humidity, outdoor air temperature and outdoor air relative humidity. The building where the air conditioning system is located is located underground. In step S20, the positive or negative value of the calculated Person correlation coefficient r represents the historical load value L T-m The indoor and outdoor characteristics and the current load value L T Positively or negatively correlated.

3. The air conditioning load prediction method based on characteristic value selection according to claim 1 or 2, characterized in that: in, Step S10 includes the following sub-steps: S11, when the enclosure structure is determined, select the historical load value L of the air conditioning system during the set time period T-m and the corresponding data of the indoor and outdoor features as a sample set; S12, using a horizontal processing method to correct the continuous outliers of the sample set: |Y(d,t)-Y(d,t-1)|>α(t), |Y(d,t)-Y(d,t+1)|>β(t), α(t) and β(t) represent the thresholds of the deviation between the historical load value at time t and the adjacent historical load values, Y(d,t) represents the historical load value at time t on the dth day, Y(d,t-1) and Y(d,t+1) represent the historical load values ​​at time t-1 and time t+1 on the dth day, respectively. S13, using the vertical processing method, based on the similar characteristics of historical load values ​​at the same time on similar days, the abnormal values ​​are replaced with data from similar dates at the same time: |Y(d,t)-a(t)|>r(t), a(t) represents the average value of the historical load value of similar days at time t, and r(t) represents the threshold between the historical load value at time t and the average value; S14: For single missing values, the average interpolation method is used to fill the gap by calculating the average of the two data before and after the missing value. To fill in two or more consecutive missing values, cubic spline interpolation is used: It is known that the function f(x) has n+1 distinct nodes on the interval [a,b], a=x0 <x1<…<x i <… <x n =The function value at b is y i =f(x i ), the constructor s(x) satisfies [x i ,x i +1] is 3, and s(x) and the first and second order derivatives are continuous on [a,b]. The calculation formula is c, d, e, and f are polynomial coefficients; S15, normalize the sample set, scale the overall data to [0,1], reduce the order of magnitude dimension gap between different data types, and eliminate abnormal results: Y Normalization Represents the normalized data, Y i represents the i-th data, Y max , Y min Respectively represent the maximum and minimum values ​​of all data; S16, standardize the sample set and transform the data into a normal distribution with a mean of 0 and a standard deviation of 1, so as to better map the relationship between the physical quantity to be predicted and the input features: Y Standardization represents the data after standardization, μ and σ represent the mean and variance of the data respectively.

4. The air conditioning load prediction method based on characteristic value selection according to claim 1 or 2, characterized in that: in, Step S20 includes the following sub-steps: S21, using the Person correlation coefficient r to evaluate the historical load value and the indoor and outdoor characteristics and the current load value L T Relevance: N represents the number of feature samples, α j and β j Respectively represent the characteristic variables between the j-th historical load value and the influencing factor, the value range of r is [-1,1], r in the range of [-1,0] indicates that the two variables are linearly negatively correlated, r in the range of [0,1] indicates that the two variables are linearly positively correlated, |r| tends to 1, indicating that the correlation between the two variables is high, and |r| tends to 0, indicating that the correlation is low; S22, respectively calculating the correlation coefficients r between several historical load values ​​and / or the indoor and outdoor features at past historical moments and the current load value, and removing the historical load values ​​and / or the indoor and outdoor features in the sample set corresponding to the calculated correlation coefficients r that violate physical principles.

5. The air conditioning load prediction method based on characteristic value selection according to claim 1 or 2, characterized in that: in, Step S30 includes the following sub-steps: S31, based on Shannon's information entropy theory, using a mutual information calculation formula to calculate the mutual information values ​​of the historical load value and the indoor and outdoor features in the data set with respect to the current load value: G represents the number of information events, u and v represent the uth and vth information events respectively, P(X u ,Y v ) indicates the Xth u ,Y v The joint probability of the occurrence of information events, P(X u ) and P(Y v ) represent the Xth u and Y v The probability of an information event occurring is measured by the mutual information value, which is used to quantitatively describe the correlation between two event sets. The larger the mutual information value, the greater the degree of mutual dependence between the two variables, and the stronger the correlation. S32, merging the mutual information values ​​calculated in step S31 and arranging them in descending order, and selecting the historical load values ​​and / or the indoor and outdoor features corresponding to all mutual information values ​​not less than the set value as feature values ​​and arranging them in descending order to form a candidate pool.

6. The air conditioning load prediction method based on characteristic value selection according to claim 5 is characterized in that: in, In step S32, the set value is 0.

185.

7. The air conditioning load prediction method based on characteristic value selection according to claim 1 is characterized in that: in, In step S40, the air conditioning load prediction model is constructed by a long short-term memory network or a BP neural network.

8. The air conditioning load prediction method based on characteristic value selection according to claim 1 is characterized in that: in, In steps S50 to S60, the prediction accuracy is calculated as follows: Correlation coefficient Mean absolute error Root mean square error Mean absolute percentage error For the predicted value The covariance with the actual value γ, and Var[γ] are the predicted values The variance of the actual value γ, n is the prediction duration, and γ z are the predicted value and actual value at the zth moment respectively. The smaller the value of ξ, the lower the prediction accuracy. The larger the values ​​of MAE, RMSE and MAPE, the lower the prediction accuracy.

9. The air conditioning load prediction method based on characteristic value selection according to claim 1 or 8, characterized in that: in, In step S50 to step S60, when the characteristic value of the same type is added as the input characteristic value to train and predict the air conditioning load prediction model, and the prediction accuracy decreases, the increase is stopped, otherwise the increase is continued until the prediction accuracy begins to decrease. When the characteristic values ​​of the same type are added as input characteristic values ​​to train and predict the air-conditioning load prediction model until the prediction accuracy begins to decrease, the characteristic values ​​of different types are started to be added as input characteristic values ​​to train and predict the air-conditioning load prediction model. If the prediction accuracy improves, the increase continues, otherwise the increase stops.

Images