A chiller load forecasting method driven by data and model fusion
Through the data and model fusion-driven method, combined with CVA and ELM, the problem of low load prediction accuracy of chiller units is solved, and efficient and accurate load prediction is achieved, which is suitable for energy resource management.
Patent Information
- Application Number
- CN202410569004.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-09
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-05-09
AI Technical Summary
The load prediction accuracy of chiller units is not high, and the existing methods have problems of insufficient efficiency and accuracy in multivariate prediction.
Using a data-model fusion-driven method, combined with canonical variable analysis (CVA) and extreme learning machine (ELM), load prediction is performed by constructing CVs-ELM models and CVA state space models, using singular value decomposition and least squares recursive estimation. , and the results are fused through the average absolute percentage error strategy.
It significantly improves the accuracy of load prediction, reduces the complexity of prediction operations, realizes the rapid and effective extraction of multivariate information, is suitable for energy resource management, and improves the accuracy and efficiency of prediction.
Smart Images

Figure CN118297239B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of chiller load prediction, and in particular to a chiller load prediction method driven by data and model fusion. Background Art
[0002] With the continuous development of science and technology, the energy consumption of large public buildings has increased year by year. Central air conditioning is a necessary equipment for large public buildings and is also a major energy-consuming equipment. Among them, chillers are the main component of central air conditioning and a large energy consumer. Therefore, the load forecasting of chillers is particularly important for energy resource management, especially in the face of complex operating conditions and load forecasting involving multiple variables. Accurate forecasting plays a great role in formulating demand scheduling and can save a lot of resources. Therefore, studying a load forecasting method driven by data and model fusion is helpful to improve the forecasting accuracy.
[0003] Canonical Variate Analysis (CVA) is a data-driven multivariate analysis method that aims to deal with the autocorrelation and correlation between variables in multivariate data. By analyzing the correlation between variables, the data dimension can be effectively reduced. It is suitable for reducing the dimension of data before extreme learning machine load forecasting. Subspace Identification Methods (SIMs) use input and output observations to directly determine the model order and state sequence of the dynamic system, thereby obtaining the state space matrix and building a model. It is mainly used to identify high-dimensional process models with a large number of inputs, outputs and states. Among the subspace identification methods, CVA has superior performance and accuracy. Recent studies have shown that Juricek and Larimore et al. have confirmed that CVA has better performance than the numerical algorithm for subspace state space system identification (N4SID) model, and is superior to the recursive model based on least squares and constrained classification regression. In addition, they also compared the results of CVA, N4SID and ARX (Auto-Regressive with Exogenous) models in the identification of chemical process dynamic models, pointing out that CVA can provide more accurate identification results and can improve the prediction accuracy to a certain extent. Therefore, modeling can be carried out based on CVA in the subspace identification method to establish a state space model for load forecasting.
[0004] Extreme Learning Machine (ELM) is a single hidden layer feedforward neural network model. Compared with the traditional feedforward neural network, the input weight matrix and the bias matrix of the hidden layer of this network are randomly generated, and the output weight is adjusted by the least squares method. It solves the problem with less execution time, fast learning speed, good generalization performance, simple structure and easy implementation, and can achieve global optimization. In 2022, Wang Qiuqiang and others invented a patent for a building air conditioning cooling load prediction method, system and product, with the patent number (CN202211574119.0). The patent proposes to determine the structural parameters of the ELM network model with the main influencing factors of the building air conditioning cooling load, and then use the Tianniu group algorithm to optimize and output the prediction result method, in which the main influencing factors are determined by principal component analysis. This method retains the integrity of the input information while improving the calculation rate, and has strong adaptability to environmental changes, stable prediction effect and high generalization ability, but the optimization algorithm makes it converge slowly, the number of iterations is large, the computational load is large, and the prediction results are inconsistent each time, which is not suitable for online prediction. Summary of the invention
[0005] The present invention utilizes open source chiller data for load forecasting. Aiming at the technical problem that the load forecasting accuracy of chillers is not high, a chiller load forecasting method driven by data and model fusion is provided. The method is divided into three parts: after completing the CVs-ELM model part and the CVA state space model part respectively, the fusion-driven load forecasting result is output according to the mean absolute percentage error strategy of each prediction result, and then compared with the real data to obtain a prediction accuracy within an acceptable range.
[0006] The inventive idea of the present invention is: the present invention provides a method for chiller load prediction driven by data and model fusion, and the method is divided into three parts: a CVs-ELM model part, in which data is divided into past data and future data, a combination matrix of the past and the future is constructed and the mutual covariance matrix is calculated, a singular value decomposition is performed on the Hankel matrix to obtain a conversion matrix, multivariate data is extracted into a small number of canonical variables CVs, the integrity of the data is retained, an extreme learning machine ELM is trained, and a CVs-ELM model is built; a CVA state space model part, after singular value decomposition is performed on the Hankel matrix, a canonical variable analysis CVA based on a subspace identification method is further used to divide the canonical variable space into a main space and a residual space, a linear combination of past observation vectors is defined according to the residual space to obtain a state vector, and then a model coefficient matrix is estimated through least squares recursion to obtain a CVA state space model; a fusion driven prediction part, input data is sent to the CVs-ELM model and the CVA state space model, and a mean absolute percentage error selection strategy is adopted for each prediction result to complete the fusion driven load prediction. The present invention integrates canonical variable analysis and extreme learning machine to complete load forecasting, realizes multi-variable and multi-model fusion-driven forecasting, and improves the accuracy of load forecasting. This is an effective exploration and supplement to the load forecasting method. In addition, in terms of energy resource management, the present invention shows broad application prospects and is expected to provide strong support for the sustainable development of the energy industry.
[0007] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: a method for driving load forecasting by data and model fusion, comprising the following steps:
[0008] Step S1: Segment the data collected within a period of time into past data y p(r) and future data f(r) , determine the length of past observations p and the length of future observations f, and set the initial value of the order n of the predicted system model.
[0009] Step S2: Construct the past and future combination matrix Y p , Y f , calculate the covariance matrix ∑ of past and future observations pp ,∑ ff And the cross-covariance matrix ∑ fp .
[0010] Step S3: Construct the Hankel matrix, perform singular value decomposition on it, and obtain the canonical variable Z based on past measurement values r And the transformation matrix J.
[0011] Step S4: Project the transformation matrix J onto the original data u to obtain the input data u' after dimensionality reduction.
[0012] Step S5: Divide the data x' into training data p 1 ,t 1 And the test data p 2 ,t 2 , and perform normalization, create and train the extreme learning machine model, set the number of input layers and hidden layers N, L, randomly initialize the connection weights IW between the input layer and the hidden layer and the hidden layer bias B 1 .
[0013] Step S6: Calculate the output value tempH of the hidden layer, use the mapping function to obtain the mapping value h of the hidden layer to the sample feature, and then calculate the connection weight LW between the hidden layer and the output layer.
[0014] Step S7: Repeat step S6 for the trained model to calculate tempH 1 and h 1 , and obtain the prediction result Y of the CVs-ELM model ELM .
[0015] Step S8: Further divide the canonical variable space into the main space and the residual space according to step S3 to obtain the canonical variable Z r Expand it into an extended vector, and then define the linear combination of past observation vectors according to the residual space to obtain the state vector x r 、x r-1 , determine the system status.
[0016] Step S9: Use linear least squares recursion to estimate the model coefficient matrices A, B, C, and D, build the model, and obtain the prediction result Y of the CVA state space model. CVA .
[0017] Step S10: For the prediction results of the two models, the mean absolute percentage error (MAPE) selection strategy is adopted to select the one with the smallest error between each sample to obtain the result Y of the fusion drive load prediction. Parallel .
[0018] Step S11: Compare the three types of prediction results with the real data, and compare the prediction performance indicators: Mean Absolute Error (MAE), Mean Relative Error (MRE), and Root Mean Square Error (RMSE), and draw conclusions.
[0019] Furthermore, in step S1, the data collected within a period of time is segmented into past data yp(r) and future data f(r) , the formula is as follows, where r represents a type of ordinal number, and p and f refer to the lengths of past and future observations, respectively.
[0020]
[0021] Further, in step S2, r=p+1,p+2,…,p+N is set to construct the past and future combination matrix Y p , Y f , the formula is as follows.
[0022] Y p =[y p(p+1) y p(p+2) … y p(p+N) ] (3)
[0023] Y f =[y f(p+1) y f(p+2) … y f(p-N) ] (4)
[0024] where Y p ∈R,Y f ∈R, set l observation variables, then y p(p+1) The last element is y (1) ,y f(p+N) The last element is y (l) , that is, the maximum number of columns of the Hankel matrix is N = lp-f + 1. Calculate the covariance matrix of past and future observations and the cross-covariance matrix between them, the formula is as follows.
[0025]
[0026] Furthermore, in step S3, CVA attempts to find y p(r) and f(r) The best linear combination between to maximize the correlation between past and future observations can be achieved by performing a singular value decomposition on the Hankel matrix as follows.
[0027]
[0028] Based on past measurements, we can get the canonical variable Z r And the transformation matrix J, the formula is as follows.
[0029]
[0030] Furthermore, in step S4, the transformation matrix J is projected onto the original data u to obtain the input data u' after the dimension is reduced.
[0031] u'=uJ (11)
[0032] Further, in step S5, the data u' is divided into training data p 1 ,t 1 And the test data p 2 ,t 2 , and normalize it. The formula is as follows, where the input data P = [p 1 p 2 ], output data T = [t 1 t 2 ].
[0033] [p 1 ,minp,maxp,t 1 ,mint,maxt]=premnmx(P,T) (12)
[0034] Create an extreme learning machine, set the number of input layers and hidden layers S and L, randomly initialize the connection weights IW between the input layer and the hidden layer and the hidden layer bias B 1 , the formula is as follows.
[0035] IW=2*rand(L,S)-1 (13)
[0036] B 1 =rand(L,1) (14)
[0037] Furthermore, in step S6, the output value tempH of the hidden layer is calculated, the mapping value h of the hidden layer to the sample feature is obtained by using the mapping function, and then the connection weight LW between the hidden layer and the output layer is calculated, and the formula is as follows.
[0038] tempH=IW*p 1 +B 1 (15)
[0039]
[0040] LW=pinv(h′)*t 1 (17)
[0041] Further, in step S7, for the trained model, step S6 is repeated to calculate tempH 1 and h 1 , and obtain the prediction result Y of the CVs-ELM model ELM .
[0042] tempH 1 =IW*P+B 1 (18)
[0043]
[0044] Y ELM =h 1 ′*LW (20)
[0045] Further, in step S8, according to the conversion matrix J in step S3, the past measured values are converted to the canonical variable space and divided into: the main space and the residual space, and the first Q main singular values determine the main space x r , the remaining singular values determine the residual space r r , so the canonical variable Z r Further expanded into an extended vector, the formula is as follows, where V x Contains the first Q columns of V,
[0046]
[0047] According to the residual space, a linear combination of past observation vectors is defined to obtain the state vector x r 、x r-1 , determine the system state. The formula is as follows, where 0∈R, the matrix is full of zero elements, is the past observation vector at the r sampling time in the process of recursive model building.
[0048]
[0049] Further, in step S9, the model coefficient matrices A, B, C, and D are estimated by linear least squares recursion. The formula is as follows, where X r =x (r:r+n) , X r-1 =x (r-1:r+n-1) , Y r =y (r:r+n) , Y r-1 =y (r-1:r+n-1) , U r =u (r:r+n) ,u r is the original data.
[0050]
[0051]
[0052] x r+1 Ax r +Bu r (28)
[0053] Build a state space model and obtain the prediction result Y of the CVA state space model CVA .
[0054] YCVA =Cx r +Du r (29)
[0055] Further, in step S10, for the prediction results of the two models, the mean absolute percentage error (MAPE) selection strategy is adopted, where is the prediction result, y i is the true value. The formula is as follows.
[0056]
[0057] Calculating MAPE for the CVs-ELM model 1 , MAPE of CVA state space model 2 , use the conditional statement to select the one with the smallest MAPE value between each sample of the two models, and get the result Y of the fusion drive load forecast. Parallel .
[0058] Furthermore, in step S11, the three types of prediction results are compared with the real data, and the prediction performance indicators MAE, MRE, and RMSE are compared to draw conclusions. The formula is as follows.
[0059]
[0060] In the above method, steps S1 to S7 are for building the CVs-ELM model, steps S8 to S9 are for building the CVA state space model, and steps S10 to S11 are for the fusion drive prediction part and drawing conclusions.
[0061] Compared with the prior art, the present invention has the following beneficial effects:
[0062] 1. The load forecasting method provided by the present invention innovatively introduces the idea of fusion-driven forecasting and proposes a chiller load forecasting method driven by data and model fusion. This method significantly improves the accuracy of load forecasting, realizes effective energy management, and injects new vitality into the field of energy utilization.
[0063] 2. The present invention cleverly combines canonical variable analysis with extreme learning machines, replacing traditional methods and innovatively constructing a CVs-ELM model, which can achieve effective and rapid extraction of multivariate complex information, thereby greatly improving the efficiency and accuracy of information processing, and providing a new approach for research and application in related fields.
[0064] 3. Based on the reduction of data dimension by canonical variable analysis, this invention further builds a CVA state space model and uses the mean absolute percentage error MAPE as the key selection strategy to complete the fusion-driven load forecasting. Compared with the forecasting method relying on the optimization algorithm, it does not require an iterative process and significantly reduces the complexity of the forecasting operation. At the same time, this method effectively ensures the accuracy of load forecasting, provides strong support for the research direction of load forecasting, and demonstrates its great potential in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0065] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.
[0066] Figure 1 The figure is the overall flow chart of load forecasting of the method of the present invention.
[0067] Figure 2 The figure is a flow chart of the model building part in the method of the present invention.
[0068] Figure 3 The load forecasting result diagram of the CVs-ELM model part is built for the embodiment of the present invention.
[0069] Figure 4 A load forecasting result diagram of the CVA state space model part is built for the embodiment of the present invention.
[0070] Figure 5 This is a load prediction result diagram of the fusion drive of the present invention according to an embodiment of the present invention.
[0071] Figure 6 It is a comparison diagram of three load forecasting result diagrams of the embodiments of the present invention.
[0072] Figure 7 It is a comparison chart of two models and fusion-driven prediction according to an embodiment of the present invention.
[0073] Among them, (a) is a comparison diagram of the absolute errors between the two models and the fusion-driven prediction in the embodiment of the present invention; (b) is a comparison diagram of the relative errors between the two models and the fusion-driven prediction in the embodiment of the present invention. DETAILED DESCRIPTION
[0074] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.
[0075] Example 1
[0076] The American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) was founded in New York, USA in 1894. In 2019, ASHRAE held an energy forecasting competition, which provided open source real data on chillers, as shown in Table 1. Based on this, a chiller load forecasting method driven by data and model fusion provided in this embodiment is as follows:
[0077] Table 1
[0078]
[0079] The real data of the chiller has 8 input variables and only 1 output data, and the total number of samples is 13615. In order to intuitively express the effectiveness of the fusion-driven chiller load prediction method, this embodiment takes part of the sample data as model training test data.
[0080] The specific steps of this embodiment are as follows:
[0081] Step 1: First split the sample data u into past data y p(r) and future data f(r) , set the observation value length p and the future observation value length f, p and f values can be set arbitrarily, in this embodiment p = 1, f = 1, set the initial value of the order n of the predicted system model, in this embodiment n = 1, r represents a type of ordinal number, the first data y at the beginning of this embodiment (r-1) That is data (1) , and so on.
[0082]
[0083] Step 2: Set r = p+1, p+2, ..., p+N to construct the past and future combination matrix Y p , Y f , Y p ∈R,Y f ∈R, set l observation variables, namely y p(p+1) The last element is y (1) ,y f(p+N) The last element is y (l) , the maximum number of columns of the Hankel matrix N = lp-f+1, in this embodiment N = 8;
[0084] Y p =[y p(p+1) y p(p+2) … y p(p+N) ],Y f =[y f(p+1) y f(p+2) … y f(p+N) ];
[0085] Step 3: After calculating the covariance matrix of past and future observations and the cross-covariance matrix between them, perform singular value decomposition on the Hankel matrix.
[0086]
[0087] Step 4: Obtain the canonical variable Z according to step 3 r and the transformation matrix J; further, projecting the transformation matrix J onto the sample data u to obtain the data u′ after the dimension is reduced;
[0088] u′=uJ;
[0089] Step 5: Divide the data u′ into training data p 1 ,t 1 And the test data p 2 ,t 2 , for the input data P = [p 1 p 2 ] and output data T = [t 1 t 2 ] for normalization.
[0090] [p 1 ,minp,maxp,t 1 ,mint,maxt=premnmx(P,T);
[0091] Step 6: Create an extreme learning machine ELM, set the number of input layers S and hidden layers L. In this embodiment, S is the main space dimension 5, L is set to 100, and randomly initialize the connection weights IW between the input layer and the hidden layer and the hidden layer bias B 1 .
[0092] IW=2*rand(L,S)-1,B 1 =rand(L,1);
[0093] Step 7: Use training data p 1 ,t 1 Training, calculate the output value tempH of the hidden layer, use the mapping function to get the mapping value h of the hidden layer to the sample feature, and calculate the connection weight LW between the hidden layer and the output layer.
[0094] tempH=IW*p 1 +B 1 , LW=pinv(h′)*t 1 ;
[0095] Step 8: After testing with the test data, the CVs-ELM model is built, and the input data P is sent into the model to obtain the prediction result Y of the CVs-ELM model. ELM .
[0096] Step 9: According to the conversion matrix J in step 3, the past measured values are converted to the canonical variable space and divided into: main space and residual space. The first Q main singular values determine the main space x r , the remaining singular values determine the residual space r r , so the canonical variable Z r Further expanded into an extended vector, where V x Contains the first Q columns of V.
[0097]
[0098] Step 10: Define the linear combination of past observation vectors according to the residual space to obtain the state vector x r 、x r-1 , determine the system state, 0∈R, the matrix is full of zero elements, is the past observation vector at the r sampling time in the process of recursive model building.
[0099]
[0100] Step 11: Use linear least squares recursion to estimate the model coefficient matrices A, B, C, and D, and then build a state space model. Send the sample data u into the model to obtain the prediction result Y of the CVA state space model. CVA .
[0101] X r =x (r:r+n) , X r-1 =x (r-1:r+n-1) , Y r =y (r:r+n) , Y r-1 =y (r-1:r+n-1) , U r =u (r:r+n) ;
[0102] x r+1 =Ax r +Bu r , Y CVA =Cx r +Du r ;
[0103] Step 12: For the prediction results of the two models, the mean absolute percentage error (MAPE) selection strategy is adopted to select the model with the smallest MAPE value between each sample of the two models to obtain the result Y of the fusion driven load prediction. Parallel .
[0104] This embodiment performs simulation according to the above steps, and can obtain the load prediction accuracy of the chiller unit of multiple models within the acceptable range, as shown in Table 2.
[0105] Table 2 Prediction accuracy within 5 kWh error
[0106]
[0107] In order to compare the three load forecasting effects more intuitively, the mean absolute error (MAE), mean relative error (MRE) and root mean square error (RMSE) are used to compare the three forecasting effects, as shown in Table 3.
[0108] Table 3 Comparison of prediction results
[0109] Evaluation indicators MAE / kWh MRE / % RMSE / kWh CVs-ELM model 6.661 5.13% 3.584 CVA state space model 2.777 2.08% 0.479 Fusion drive of the present invention 1.963 1.48% 0.348
[0110] It can be seen from Tables 2 and 3 above that the fusion drive proposed in this embodiment is significantly better than the prediction results of the CVs-ELM model part and the CVA state space model part in terms of prediction results. It greatly improves the accuracy of chiller load forecasting and brings breakthroughs in the field of energy management. The innovative data and model fusion driven method adopted in this embodiment not only organically combines data and models, creates a new way to reduce dimensions, and effectively explores the potential value of multiple variables, but also further builds the state space model part for load forecasting, significantly improving the prediction accuracy. The outstanding performance of this method enables the fusion driven prediction to achieve good results. In summary, the chiller load forecasting method driven by data and model fusion in this embodiment shows high adaptability and value in energy resource forecasting.
[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A chiller load forecasting method driven by data and model fusion, characterized in that: The following steps are involved: Step S1: Segment the data collected within a period of time into past data y p(r) and future data f(r) , determine the length of past observations p and the length of future observations f, and set the initial value of the order n of the predicted system model; Step S2: Construct the past and future combination matrix Y p , Y f , calculate the covariance matrix Σ of past and future observations pp ,Σ ff and the cross-covariance matrix Σ between them fp ; Step S3: Construct the Hankel matrix, perform singular value decomposition on it, and obtain the canonical variable Z based on past measurements r and transformation matrix J; Step S4: Project the transformation matrix J onto the original data u to obtain the input data u' after the dimension is reduced; In step S4, the transformation matrix J is projected onto the original data u to obtain the input data u' after dimensionality reduction. u'=uJ (1); Step S5: Divide the data x' into training data p1, t1 and test data p2, t2, and perform normalization processing, create and train the extreme learning machine model, set the number of input layers and hidden layers N, L, and randomly initialize the connection weight IW between the input layer and the hidden layer and the hidden layer bias B1; Step S6: Calculate the output value tempH of the hidden layer, use the mapping function to obtain the mapping value h of the hidden layer to the sample feature, and then calculate the connection weight LW between the hidden layer and the output layer; Step S7: Repeat step S6 for the trained model, calculate tempH1 and h1, and obtain the prediction result Y of the CVs-ELM model ELM ; Step S8: Further divide the canonical variable space into the main space and the residual space according to step S3 to obtain the canonical variable Z r Expand it into an extended vector, and then define the linear combination of past observation vectors according to the residual space to obtain the state vector x r 、x r-1 , determine the system status; Step S9: Use linear least squares recursion to estimate the model coefficient matrices A, B, C, and D, build the model, and obtain the prediction result Y of the CVA state space model. CVA ; In step S9, the model coefficient matrices A, B, C, and D are estimated by linear least squares recursion, as follows: r =x (r:r+n) , X r-1 =x (r-1:r+n-1) , Y r =y (r:r+n) , Y r-1 =y (r-1:r+n-1) , U r =u (r:r+n) ,u r is the original data, x r+1 =Ax r +Bu r (29) Build a state space model and obtain the prediction result Y of the CVA state space model CVA ; Y CVA =Cx r +You r (30); Step S10: For the prediction results of the two models, the mean absolute percentage error selection strategy is adopted to select the smallest error between each sample and obtain the result Y of the fusion drive load prediction. Parallel ; In step S10, for the prediction results of the two models, the mean absolute percentage error (MAPE) selection strategy is adopted, where is the prediction result, y i is the true value, and the formula is as follows; Calculate the MAPE1 of the CVs-ELM model and the MAPE2 of the CVA state space model, and use the conditional statement to select the smaller MAPE value between each sample of the two models to obtain the result Y of the fusion drive load forecast Parallel ; Step S11: Compare the three types of prediction results with the actual data, compare the prediction performance indicators: mean absolute error, mean relative error, root mean square error, and draw conclusions.
2. The chiller load forecasting method driven by data and model fusion according to claim 1 is characterized in that: In step S1, the data collected within a period of time is segmented and divided into past data y p(r) and future data f(r) , the formula is as follows, where r represents a type of ordinal number, p and f refer to the length of past and future observations respectively; 3. The chiller load forecasting method driven by data and model fusion according to claim 1 is characterized in that: In step S2, set r = p+1, p+2, ..., p+N to construct the past and future combination matrix Y p , Y f , the formula is as follows: AND p =[and p(p+1) and p(p+2) …and p(p+N) ] (5) AND f =[and f(p+1) and f(p+2) …and f(p+N) ] (6) where Y p ∈R,Y f ∈R, set l observation variables, then y p(p+1) The last element is y (1) ,y f(p+N) The last element is y (l) , that is, the maximum number of columns of the Hankel matrix is N = lp-f + 1, and the covariance matrix of past and future observations and the cross-covariance matrix between them are calculated as follows:
4. The chiller load forecasting method driven by data and model fusion according to claim 1 is characterized in that: In step S3, CVA attempts to find y p(r) and f(r) The best linear combination between to maximize the correlation between past and future observations is achieved by performing a singular value decomposition on the Hankel matrix. The formula is as follows: Based on past measurements, we can get the canonical variable Z r And the transformation matrix J, the formula is as follows, 5. The chiller load forecasting method driven by data and model fusion according to claim 1 is characterized in that: In step S5, the data u' is divided into training data p1, t1 and test data p2, t2, and normalized. The formula is as follows, where input data P = [p1p2], output data T = [t1t2], [p1,minp,maxp,t1,mint,maxt]=premnmx(P,T) (13) Create an extreme learning machine, set the number of input layers and hidden layers S and L, randomly initialize the connection weights IW between the input layer and the hidden layer and the hidden layer bias B1. The formula is as follows: IW=2*rand(L,S)-1 (14) B1=rand(L,1) (15).
6. The chiller load forecasting method driven by data and model fusion according to claim 1 is characterized in that: In step S6, the output value tempH of the hidden layer is calculated, and the mapping value h of the hidden layer to the sample feature is obtained by using the mapping function, and then the connection weight LW between the hidden layer and the output layer is calculated. The formula is as follows: tempH=IW*p1+B1 (16) LW=pinv(h′)*t1 (18).
7. The chiller load forecasting method driven by data and model fusion according to claim 1 is characterized in that: In step S7, for the trained model, repeat step S6 to calculate tempH1 and h1 to obtain the prediction result Y of the CVs-ELM model. ELM , tempH1=IW*P+B1 (19) AND ELM =h1'*LW (21)。 8. The chiller load forecasting method driven by data and model fusion according to claim 1 is characterized in that: In step S8, according to the conversion matrix J in step S3, the past measured values are converted to the canonical variable space and divided into: the main space and the residual space. The first Q main singular values determine the main space x r , the remaining singular values determine the residual space r r , the canonical variable Z r Further expanded into an extended vector, the formula is as follows, where V x Contains the first Q columns of V, According to the residual space, a linear combination of past observation vectors is defined to obtain the state vector x r 、x r-1 , determine the system state, the formula is as follows, where 0∈R, the matrix is full of zero elements, is the past observation vector at the r sampling time in the process of recursive model building, 9. The chiller load forecasting method driven by data and model fusion according to claim 1 is characterized in that: In step S11, the three types of prediction results are compared with the real data, and the predicted performance indicators MAE, MRE, and RMSE are compared to draw conclusions. The formula is as follows: Steps S1 to S7 are for building the CVs-ELM model, steps S8 to S9 are for building the CVA state space model, and steps S10 to S11 are for the fusion drive prediction part and drawing conclusions.
Citation Information
Patent Citations
Building air conditioner cooling load prediction method, system and product
CN115860234A
Short-term Load Forecasting Method Based on TCN and IPSO-LSSVM Combined Model
AU2020104000A4
Multivariate detection of abnormal conditions in a process plant
CN101542407A
Ultra-short-term load prediction method of extreme learning machine
CN104299043A
Short-term electric load prediction method based on improved genetic algorithm for optimizing extreme learning machine
CN108876054A
