Building load prediction method based on forward decomposition denoising and deep integrated learning

Abstract

The invention relates to a building load prediction method based on forward decomposition denoising and deep integrated learning, and the method comprises the steps: carrying out the forward decomposition of original load data into a plurality of IMFs components through employing a CEEMDAN algorithm, removing a high-frequency noise component, and carrying out the reconstruction fusion of residual components, and obtaining clean load data; then, a DWedRVFL neural network model is constructed, the DWedRVFL neural network model captures the dynamic change of the load sequence by adopting the predicted latest precision, and a diversity index is constructed to utilize the multi-scale features of different output layers; in addition, a sorting strategy is constructed to integrate the latest precision and diversity contributions, and meanwhile, the influence of an abnormal prediction value on combined prediction is avoided; and then, the historical load, the weather condition and the personnel occupation data form a connection matrix, and the connection matrix is fed into the DWedRVFL neural network model for learning, so that the accuracy and the stability of building load prediction are improved. The method is beneficial to improving the precision and stability of short-term building load prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of load forecasting, and in particular to a building load forecasting method based on forward decomposition denoising and deep ensemble learning. Background Art

[0002] Promoting sustainable development of cities and society will become increasingly important for solving energy and environmental crises. Various methods based on building energy management systems (BEMS), such as low-carbon energy comprehensive evaluation, carbon trading mechanism, and carbon footprint assessment, have gradually been applied. However, the efficient and stable operation of BEMS requires accurate prediction of building loads.

[0003] Building load data may be affected by measurement noise and external interference during the collection process, which makes it difficult to identify effective information. In addition, factors such as human activities and weather conditions cause significant irregular fluctuations in building loads, and the time-varying nature of the loads themselves further exacerbates the complexity of load pattern identification. Therefore, in order to improve the accuracy of building load prediction, it is necessary to remove noise to improve data quality, and the strong volatility and time-varying nature of building loads should also be fully considered in the model design stage.

[0004] The Chinese patent application number 202311481491.1 discloses a short-term building load forecasting method, which extracts the load data of a building within a certain period of time as the original data, processes the original data for outliers and missing values, and obtains feature data; establishes a TCN-LSTM model based on the residual self-attention mechanism, trains the model using feature data, and obtains the predicted load value; optimizes the hyperparameters of the RSA-TCN-LSTM model using the improved sparrow search algorithm; and uses the optimized RSA-TCN-LSTM prediction model to predict the building load. This method does not take into account the noise interference in the process of building load collection, and it is difficult to obtain high-quality load data.

[0005] The Chinese patent application number 202310740982.7 discloses a commercial building load forecasting method and system based on an informer network. The method collects historical energy consumption metadata information of different types of commercial buildings to construct load training data sets corresponding to different types of commercial buildings; obtains temperature data from the meteorological bureau to obtain a temperature data set; trains the temperature data set and different types of load data based on the informer network to obtain load forecasting models and temperature forecasting models corresponding to different types of commercial buildings; combines the temperature data predicted by the temperature forecasting model and the load forecasting model to obtain the final load forecasting result. This method does not customize the modeling design for the strong volatility and time-varying characteristics of building loads, which affects the load forecasting accuracy. Summary of the invention

[0006] The purpose of the present invention is to provide a building load forecasting method based on forward decomposition denoising and deep ensemble learning, which is conducive to improving the accuracy and stability of short-term building load forecasting.

[0007] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a building load forecasting method based on forward decomposition denoising and deep integrated learning. For the original load data, the CEEMDAN algorithm is first used to forward decompose the load data into multiple IMFs components, and the remaining components are reconstructed and fused after removing the high-frequency noise components to obtain clean load data; then, in order to cope with the strong volatility and time-varying nature of the building load data, a DWedRVFL neural network model is constructed, and the DWedRVFL neural network model uses the latest prediction accuracy to capture the dynamic changes of the load sequence, and constructs a diversity index to utilize the multi-scale features of different output layers; in addition, a sorting strategy is constructed to integrate the contribution of the latest accuracy and diversity, while avoiding the influence of abnormal prediction values ​​on the combined prediction; then, the connection matrix composed of historical load, weather conditions and personnel occupancy data is fed into the DWedRVFL neural network model for learning, thereby improving the accuracy and stability of building load forecasting.

[0008] Furthermore, the implementation method of the CEEMDAN algorithm is:

[0009] Step 1: Assume that the collected original load data is The forward rolling decomposition method is used instead of the decomposition of the entire signal. That is, CEEMDAN decomposes the data within a forward rolling window size w. The data within w only includes known load data and does not involve future data. The load data used for decomposition is expressed as

[0010] Step 2: Add Gaussian white noise sequence; in the original load data S L Add a set of Gaussian white noise sequences to form new data S new , as shown below:

[0011] S new =S L +γ0G i

[0012] Where i is the number of trials; G i is a normally distributed Gaussian noise sequence; γ0 is the ratio of data to noise, which is used to control the additional noise of the original data set;

[0013] Step 3: According to the EMD method, S new Decompose; from the signal Snew Extract the local maximum and minimum points, and calculate the mean M of the upper envelope U and the lower envelope L according to the local extreme points:

[0014]

[0015] Step 4: From the original signal S new Subtract the mean M from the modal component h1 to obtain the preliminary modal characteristic component h1:

[0016] h1=S new -M

[0017] Step 5: Check whether h1 satisfies the two conditions that the number of zero crossing points should be equal to the number of extreme points and the amplitude is symmetric; if not, it is regarded as a new signal, and steps 3 and 4 are repeated until the IMF conditions are met. After convergence, the first component of EMD is initially obtained; then the first component of CEEMDAN, IMF1, is obtained by calculating the mean, as shown below:

[0018]

[0019] Where T represents the total number of times white noise is added;

[0020] Step 6: Calculate the first residual term r1 as follows:

[0021] r1=S L -IMF1

[0022] Step 7: Add noise to the first residual term r1 to obtain a new sequence r1+γ1E1(G i ), the new sequence is decomposed as follows to obtain the second component IMF2 and the second residual term r2:

[0023]

[0024] r2=r1-IMF2

[0025] Step 8: Calculate the k-order residual term r based on the same method as step 7 k and the k+1th order component IMF k+1 :

[0026] r k =r k-1 -IMF k

[0027]

[0028] Step 9: Repeat step 8 until the residual term can no longer be decomposed. The final residual term R is calculated as follows:

[0029]

[0030] Step 10: Assuming that component IMF1 is the high-frequency noise component that needs to be removed, the remaining components except IMF1 and the residual term R are reconstructed into clean load data:

[0031]

[0032] Step 11: X L represents the denoised historical load sequence, X n and X w They represent the building occupancy data obtained from the punch card system and the weather forecast data obtained from the BEMS system respectively; the prediction modeling goal is to establish the following mapping relationship:

[0033]

[0034] In the formula, is the predicted value at the next moment, e is the error; the rolling window w determines the amount of historical load data used; among them, the historical load series is expressed as And X n , X W is a scalar; [X L ,X n ,X W ] forms a connection matrix X as the input of the prediction model, and the output is the predicted load value of the next time step.

[0035] Furthermore, the implementation method of the DWedRVFL neural network model is:

[0036] The input data is Where N and m are the number of samples and features of the input data; P is defined as the number of neurons in the enhancement layer; the feature representation of the first enhancement layer of DWedRVFL is Calculated by the following formula:

[0037] H 1 =g(XB1)

[0038] Among them, g(·) represents the nonlinear activation function, Randomly generated connection weights for the first enhancement layer;

[0039] For the DWedRVFL model with L enhancement layers, the input of the lth enhancement layer consists of the initial input feature X and the output H of the previous enhancement layer. l-1 The calculation of the lth enhancement layer can be defined as follows:

[0040]

[0041] in, is the weight and enhancement feature of the l-th enhancement layer, is the output of the lth layer; then the loss function of the lth enhancement layer can be expressed as:

[0042]

[0043] In the formula, β l is the connection weight of the lth layer, λ is the regularization parameter; generally speaking, β l The calculation of can be calculated analytically using the Moore-Penrose pseudo-inverse method or the ridge regression method, expressed as:

[0044]

[0045] In the formula, is the extended input matrix, I is the identity matrix; each RVFL module in DWedRVFL corresponds to a decision subtask, and the output weights of each layer are calculated separately; all weights of the hidden layers of DWedRVFL are randomly generated before the start of training and remain fixed during the network training process;

[0046] In view of the high volatility and time-varying characteristics of building loads, a dynamic integration module is designed. The outputs of different enhancement layers are combined by dynamically updating weights to achieve a trade-off between the latest accuracy and diversity. The predicted output of the network with L enhancement layers is The design goal of the integration module is to find a set of combined weights that minimize the root mean square error (RMSE) between the predicted value and the actual value. The objective function is expressed as:

[0047]

[0048] Furthermore, in order to enhance the adaptability of DWedRVFL to the time-varying characteristics of building loads, DWedRVFL introduces a contribution index based on the latest accuracy, which is defined as a function of the latest measurement prediction error:

[0049]

[0050] In the formula, f l It represents the contribution index of the lth layer based on the latest accuracy at time t-1; in order to prevent the weight of the dominant output layer from being too large, the contribution index values ​​of each layer are sorted from large to small using the sorting method, and the following order is obtained:

[0051] f1≥f2≥…≥f l ≥…≥f L

[0052] Then, the latest accuracy contribution of layer l is sorted as:

[0053]

[0054] in, represents the index value after sorting at the lth level, F l Indicates the ranking of its contribution to the latest accuracy; if the latest accuracy contribution of the lth layer is the largest, its ranking index value is 1, E l equal to L; for each output layer, the ranking value F l The higher it is, the greater its contribution to the latest accuracy, which means it has a higher level of accuracy.

[0055] Furthermore, DWedRVFL introduces diversity to ensure the effective utilization of multi-scale enhanced features from different output layers, thereby effectively coping with the volatility of building loads; DWedRVFL promotes the diversity of overall predictions by giving greater weight to the output layers with diversity contributions; first, the Euclidean distance between each layer's prediction and the predictions of other layers is defined as the diversity contribution:

[0056]

[0057] Among them, E l is the sum of the Euclidean distances between the prediction of the lth layer and the predictions of other layers at time t. According to the prediction distribution, E l Defined as the diversity contribution index; then sort the diversity contribution index in descending order, and the order obtained is as follows:

[0058] E1≥E2≥…≥E l ≥…≥E L

[0059] Then the diversity ranking of the lth layer is assigned as:

[0060]

[0061] In the formula, For E l The index after sorting in descending order, is the diversity ranking based on the predicted distribution; for the output layer l, the greater its diversity contribution, The bigger;

[0062] In order to obtain a more comprehensive diversity metric, DWedRVFL further defines diversity based on prediction performance. The diversity contribution index based on prediction performance is calculated as the difference between a single prediction and the median of all predictions. When calculating the deviation, the true label at time t is unknown. Therefore, the latest prediction error is used to calculate the deviation. The deviation of the lth layer at time (t-1) is for:

[0063]

[0064] Will Sort in descending order to get the ranking of the lth layer:

[0065]

[0066] in, are the index based on prediction performance and the diversity contribution ranking respectively; according to the above two diversity calculation results, the overall diversity contribution ranking of the lth layer is defined as:

[0067]

[0068] Among them, d l is the overall diversity contribution indicator for layer l, is d l Index value after reverse sorting; D l is the ranking value of the lth layer based on the comprehensive diversity index;

[0069] Comprehensive F l and D l The sorting information defines the final contribution index r of the lth output layer l for:

[0070] r l =αF l +(1-α)D l

[0071] Among them, α is the trade-off parameter;

[0072] DWedRVFL uses a ranking-based strategy to calculate the final ranking R of the lth layer l and combination weight

[0073] r1≥r2≥…≥r l ≥…≥r L ,

[0074]

[0075] in, For r l Index value after descending order;

[0076] Calculate the combined prediction value of DWedRVFL at time t according to the sorting strategy The calculation method is:

[0077]

[0078] Compared with the prior art, the present invention has the following beneficial effects: first, the present invention adopts the forward CEEMDAN algorithm to forward decompose and denoise the original load data, thereby obtaining high-quality load data, which is helpful for the subsequent machine learning process; secondly, the present invention fully considers the strong volatility and time-varying characteristics of building load in the prediction modeling stage; proposes a diversity metric to make full use of the multi-scale characteristics of different output layers to obtain a more stable prediction; develops a dynamic integration weight based on the latest accuracy to adapt to the time-varying nature of the load; designs a sorting strategy to integrate the contribution of diversity and the latest accuracy to the combined prediction, while avoiding the adverse effects of outliers. Compared with patent 202311481491.1, the present invention considers the noise interference in the load data collection process, and obtains higher quality load data after denoising by the forward CEEMDAN algorithm; compared with patent 202310740982.7, the present invention takes into account the volatility and time-varying nature of the building load in the load prediction modeling stage, designs the DWedRVFL prediction network to accurately identify the change pattern of the building load, and improves the accuracy and stability of short-term building load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 is a flow chart of a method implementation of an embodiment of the present invention;

[0080] Figure 2 is a flow chart of the implementation of the CEEMDAN algorithm in an embodiment of the present invention;

[0081] Figure 3 It is a training framework diagram of the DWedRVFL neural network model in an embodiment of the present invention;

[0082] Figure 4 is a graph of the original load data collected on a working day in an embodiment of the present invention;

[0083] Figure 5 is a scrolling data graph of the first window in an embodiment of the present invention;

[0084] Figure 6 is a data graph after adding noise in an embodiment of the present invention;

[0085] Figure 7 is a data diagram of the mean value M of the upper envelope U and the lower envelope L in an embodiment of the present invention;

[0086] Figure 8 is a preliminary modal characteristic component data diagram obtained in an embodiment of the present invention;

[0087] Fig. 9 is a data diagram of the first component IMF1 and the first residual term r1 obtained in an embodiment of the present invention;

[0088] Fig.10 is a data diagram of the second component IMF2 and the second residual term r2 obtained in an embodiment of the present invention;

[0089] Fig.11 It is a data diagram of all IMFs and the final remaining residual term R obtained in the embodiment of the present invention;

[0090] Fig.12 is a first window denoised load data graph obtained in an embodiment of the present invention;

[0091] Fig.13 It is a prediction result diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0092] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0093] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0094] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0095] The measurement noise in the process of building load data collection and the strong volatility and time-varying nature of building loads lead to a decrease in prediction accuracy. To solve the above problems, this embodiment provides a building load prediction method based on forward decomposition denoising and deep ensemble learning. For the original load data, the CEEMDAN algorithm is first used to forward decompose the load data into multiple IMFs components, and the remaining components are reconstructed and fused after removing the high-frequency noise components to obtain clean load data; then, in order to cope with the strong volatility and time-varying nature of building load data, a dynamic weight ensemble deep random vector functional link (DWedRVFL) neural network model is constructed. The DWedRVFL neural network model uses the latest prediction accuracy to capture the dynamic changes of the load sequence, and constructs a diversity index to utilize the multi-scale features of different output layers; in addition, a sorting strategy is constructed to integrate the contribution of the latest accuracy and diversity, while avoiding the influence of abnormal prediction values ​​on the combined prediction; then, the connection matrix composed of historical load, weather conditions and occupancy data is fed into the DWedRVFL neural network model for learning, thereby improving the accuracy and stability of building load prediction.

[0096] Figure 1 is a flowchart for implementing the method of this embodiment. This method is applicable to general buildings that can be used for residential, commercial or educational purposes; the building has energy-consuming equipment and generates loads; the load data is recorded and stored by the load metering equipment in the building. This method is not applicable to old buildings that have not been electrified and are not equipped with load metering equipment.

[0097] Figure 2 : is a flow chart of the implementation of the CEEMDAN algorithm in this embodiment. Figure 2 As shown, the implementation method of the CEEMDAN algorithm is:

[0098] Step 1: Assume that the collected original load data is Since the CEEMDAN decomposition process involves the information of the entire data series, it incorrectly decomposes the entire time series S L This may lead to information leakage. Therefore, this paper adopts the forward rolling decomposition method instead of the decomposition of the entire signal, that is, CEEMDAN decomposes the data within a forward rolling window size w, and the data within w only includes known load data and does not involve future data; the load data used for decomposition can be expressed as

[0099] Step 2: Add Gaussian white noise sequence. L Add a set of Gaussian white noise sequences to form new data Snew , as shown below:

[0100] S new =S L +γ0G i

[0101] Where i is the number of trials; G i is a normally distributed Gaussian noise sequence; γ0 is the ratio of data to noise, which is used to control the additional noise of the original data set.

[0102] Step 3: First, use the Empirical Mode Decomposition (EMD) method to decompose S new First, from the signal S new Extract the local maximum and minimum points, and calculate the mean M of the upper envelope U and the lower envelope L according to the local extreme points:

[0103]

[0104] Step 4: From the original signal S new Subtract the mean M from the modal component h1 to obtain the preliminary modal characteristic component h1:

[0105] h1=S new -M

[0106] Step 5: Check whether h1 satisfies the two conditions that the number of zero crossing points should be equal to the number of extreme points and the amplitude is symmetric; if not, it is regarded as a new signal, and steps 3 and 4 are repeated until the IMF conditions are met. After convergence, the first component of EMD is initially obtained. Then, the first component of CEEMDAN, IMF1, is obtained by calculating the mean, as shown below:

[0107]

[0108] Where T represents the total number of times white noise is added.

[0109] Step 6: Calculate the first residual term r1 as follows:

[0110] r1=S L -IMF1

[0111] Step 7: Add noise to the first residual term r1 to obtain a new sequence r1+γ1E1(G i ), the new sequence is decomposed as follows to obtain the second component IMF2 and the second residual term r2:

[0112]

[0113] r2=r1-IMF2

[0114] Step 8: Calculate the k-order residual term r based on the same method as step 7 k and the k+1th order component IMF k+1 :

[0115] r k =r k-1 -IMF k

[0116]

[0117] Step 9: Repeat step 8 until the residual term can no longer be decomposed. The final residual term R is calculated as follows:

[0118]

[0119] Step 10: Assuming that component IMF1 is the high-frequency noise component that needs to be removed, the remaining components except IMF1 and the residual term R are reconstructed into clean load data:

[0120]

[0121] Step 11: X L represents the denoised historical load sequence, X n and X w They represent the building occupancy data obtained from the punch card system and the weather forecast data obtained from the BEMS system. The prediction modeling goal is to establish the following mapping relationship:

[0122]

[0123] In the formula, is the predicted value at the next moment, and e is the error. The rolling window w determines the amount of historical load data used. Among them, the historical load series is expressed as And X n , X W is a scalar; [X L ,X n ,X W ] forms a connection matrix X as the input of the prediction model, and the output is the predicted load value of the next time step.

[0124] Figure 3 : is a training framework diagram of the DWedRVFL neural network model in this embodiment. Figure 3 As shown, the implementation method of the DWedRVFL neural network model is:

[0125] The input data is Where N and m are the number of samples and features of the input data. Where P is defined as the number of neurons in the enhancement layer. The feature representation of the first enhancement layer of DWedRVFL is It can be calculated by the following formula:

[0126] H 1 =g(XB1)

[0127] Where g(·) represents a nonlinear activation function, such as sigmoid, tanh or relu, Randomly generated connection weights for the first enhancement layer.

[0128] For the DWedRVFL model with L enhancement layers, the input of the lth enhancement layer consists of the initial input feature X and the output H of the previous enhancement layer. l-1 The calculation of the lth enhancement layer can be defined as follows:

[0129] H l =g([H l-1 ,X]B l )

[0130]

[0131] in, is the weight and enhancement feature of the l-th enhancement layer, is the output of the lth layer. Then the loss function of the lth enhancement layer can be expressed as:

[0132]

[0133] In the formula, β l is the connection weight of the lth layer, and λ is the regularization parameter. Generally speaking, β l The calculation of can be calculated analytically using the Moore-Penrose pseudo-inverse method or the ridge regression method, expressed as:

[0134]

[0135] In the formula, is the expanded input matrix, and I is the identity matrix. Each RVFL module in DWedRVFL corresponds to a decision subtask, and the output weights of each layer are calculated separately. All weights of the hidden layers of DWedRVFL are randomly generated before training begins and remain fixed during network training.

[0136] In view of the high volatility and time-varying characteristics of building loads, a dynamic integration module is designed, and the outputs of different enhancement layers are combined by dynamically updating weights to achieve a trade-off between the latest accuracy and diversity. The predicted output of the network with L enhancement layers is The design goal of the integration module is to find a set of combined weights that minimize the root mean square error (RMSE) between the predicted value and the actual value. The objective function can be expressed as:

[0137]

[0138] In order to enhance the adaptability of DWedRVFL to the time-varying characteristics of building loads, DWedRVFL introduces a contribution index based on the latest accuracy, which is defined as a function of the latest measurement prediction error:

[0139]

[0140] In the formula, f l It represents the contribution index of the lth layer based on the latest accuracy at time t-1; in order to prevent the weight of the dominant output layer from being too large, the contribution index values ​​of each layer are sorted from large to small using the sorting method, and the following order is obtained:

[0141] f1≥f2≥…≥f l ≥…≥f L

[0142] Then, the latest accuracy contribution of layer l is sorted as:

[0143]

[0144] in, represents the index value after sorting at the lth level, F l Indicates the ranking of its contribution to the latest accuracy. For example, if the latest accuracy contribution of layer l is the largest, its ranking index value is 1, and F l is equal to L. For each output layer, the ranking value F l The higher it is, the greater its contribution to the latest accuracy, which means it has a higher level of accuracy.

[0145] Considering the nonlinear fluctuation of building load, single-layer prediction may produce unstable predictions. Diversity is introduced to ensure the effective utilization of multi-scale enhanced features from different output layers, thereby effectively coping with the volatility of building load. DWedRVFL promotes the diversity of overall prediction by giving greater weight to the output layers with diversity contributions. First, the Euclidean distance between each layer prediction and the predictions of other layers is defined as the diversity contribution:

[0146]

[0147] Among them, E l is the sum of the Euclidean distances between the prediction of the lth layer and the predictions of other layers at time t. According to the prediction distribution, E l Defined as the diversity contribution index; then sort the diversity contribution index in descending order, and the order obtained is as follows:

[0148] E1≥E2≥…≥E l ≥…≥E L

[0149] Then the diversity ranking of the lth layer is assigned as:

[0150]

[0151] In the formula, For E l The index after sorting in descending order, is the diversity ranking based on the predicted distribution. For the output layer l, the greater its diversity contribution, The bigger.

[0152] In order to obtain a more comprehensive diversity metric, DWedRVFL further defines diversity based on prediction performance. The diversity contribution index based on prediction performance is calculated as the difference between a single prediction and the median of all predictions. When calculating the deviation, the true label at time t is unknown. Therefore, we use the latest prediction error to calculate the deviation. Then the deviation of the lth layer at time (t-1) is for:

[0153]

[0154] Will Sort in descending order to get the ranking of the lth layer:

[0155]

[0156] in, are the index based on prediction performance and the diversity contribution ranking respectively. Based on the above two diversity calculation results, the overall diversity contribution ranking of the lth layer is defined as:

[0157]

[0158] Among them, d l is the overall diversity contribution indicator for layer l, is d l The index value after sorting in reverse order. D l is the ranking value of the lth layer based on the comprehensive diversity index.

[0159] Comprehensive F l and D l The sorting information defines the final contribution index r of the lth output layer l for:

[0160] r l =αF l +(1-α)D l

[0161] Among them, α is a trade-off parameter with a value range of 0.1-0.9, which can be adjusted according to different data sets.

[0162] DWedRVFL uses a ranking-based strategy to calculate the final ranking R of the lth layer l and combination weight

[0163] r1≥r2≥…≥r l ≥…≥r L ,

[0164]

[0165] in, For r l The index value after descending order.

[0166] Calculate the combined prediction value of DWedRVFL at time t according to the sorting strategy The calculation method is:

[0167]

[0168] In summary, the ranking-based dynamic integration strategy adopted by DWedRVFL provides a balanced combination weight allocation scheme, which ensures that excellent predictions obtain greater weights while reducing the impact of outlier predictions, thereby producing a more reasonable combination prediction value.

[0169] The specific implementation of the present invention is described in detail below with reference to data examples. The specific steps are as follows:

[0170] Step 1: The original load data S collected on working days of a building in June 2023 L The load curve and basic statistical indicators are as follows: Figure 4 As shown in Table 1. The forward rolling decomposition method is used instead of the decomposition of the entire signal, that is, CEEMDAN decomposes the data within a forward rolling window size of w = 48, and the data within w only includes known load data and does not involve future data. Then the 48 load data S used for decomposition in the first rolling window L like Figure 5 As shown:

[0171] Table 1

[0172]

[0173] Step 2: Add Gaussian white noise sequence. L Add a set of Gaussian white noise sequences to form new data S new , calculated as follows:

[0174] S new =S L +γ0G i

[0175] Where i is the number of trials, which is set to 3; G i is a Gaussian noise sequence with normal distribution N(0,0.5); γ0=0.6 is the ratio of data to noise, which is used to control the additional noise of the original data set; then S new like Figure 6 shown.

[0176] Step 3: First, use the EMD method to calculate S new First, from the signal S new The local maximum and minimum points are extracted from the local extreme points, and the mean M of the upper envelope U and the lower envelope L are calculated according to the local extreme points, as shown in Figure 7 As shown:

[0177]

[0178] Step 4: From the original signal S new Subtract the mean M from the modal component h1 to obtain the preliminary modal characteristic component h1. Figure 8 Said:

[0179] h1=S new -M

[0180] Step 5: Check whether h1 satisfies the two conditions that the number of zero crossing points should be equal to the number of extreme points and the amplitude is symmetric; if not, it is regarded as a new signal, and steps 3 and 4 are repeated until the IMF conditions are met. After convergence, the first component of EMD is initially obtained. Then, the first component of CEEMDAN, IMF1, is obtained by calculating the mean, as shown below:

[0181]

[0182] Wherein, T=3 is the total number of times white noise is added.

[0183] Step 6: Calculate the first residual term r1 as follows:

[0184] r1=S L -IMF1

[0185] Then the first component IMF1 and the first residual term r1 are obtained as Fig. 9 shown.

[0186] Step 7: Add noise to the first residual term r1 to obtain a new sequence r1+γ1E1(G i), the new sequence is decomposed according to the following formula to obtain the second component IMF2 and the second residual term r2. The result is as follows Fig.10 As shown:

[0187]

[0188] r2=r1-IMF2

[0189] Step 8: Calculate the k-order residual term r based on the same method as step 7 k and the k+1th order component IMF k+1 :

[0190] r k =r k-1 -IMF k

[0191]

[0192] Step 9: Repeat step 8 until the residual term can no longer be decomposed. The final residual term R is calculated as follows:

[0193]

[0194] All IMFs and the final residual term R are calculated as follows: Fig.11 shown.

[0195] Step 10: Assuming that component IMF1 is the high-frequency noise component that needs to be removed, the remaining components except IMF1 and the residual term R are reconstructed into clean load data according to the following formula:

[0196]

[0197] The denoised load data obtained in the first window is as follows Fig.12 shown.

[0198] Step 11: Construct a data set for model training. Continue to denoise the load data in the window according to steps 1-10 until the complete denoised historical load sequence X is obtained. L Next, the number of building occupants X obtained from the building clock-in system n and the weather forecast variable X obtained from the BEMS system w ; Since the outdoor average temperature is highly correlated with the load, the X w Represents the outdoor temperature variable. The prediction modeling goal is to establish the following mapping relationship:

[0199]

[0200] In the formula, is the predicted value at the next moment, and e is the error. The rolling window w=48 is the amount of historical load data used. The historical load sequence is expressed as And X n , X W Is a scalar. The connection matrix X = [X L ,X n ,X W ] is used as the input data set of the prediction model, with a feature dimension of 48+1+1=50 and a sample size of 1008.

[0201] Step 12: Construct the DWedRVFL neural network model. The input data is Where N, m are the number of samples and features of the input data, which are 1008 and 50 respectively. The training set, validation set and test set are divided into 7:1:2 ratios, and the N, m of the training data are 705 and 50 respectively. Where P = 160 is defined as the number of neurons in the enhancement layer. The training parameters of DWedRVFL are shown in Table 2. The feature representation of the first enhancement layer of DWedRVFL is It can be calculated by the following formula:

[0202] H 1 =g(XB1),

[0203] Where g(·) represents a nonlinear activation function, and sigmoid is used in this paper. Randomly generated connection weights for the first enhancement layer.

[0204] Table 2

[0205]

[0206]

[0207] Step 13: For the DWedRVFL model with L = 10 enhancement layers, the input of the lth enhancement layer consists of the initial input feature X and the output H of the previous enhancement layer l-1 The composition is defined according to the following formula:

[0208] H l =g([H l-1 ,X]B l ),

[0209]

[0210] in is the weight and enhancement feature of the l-th enhancement layer, is the output of the lth layer. Then the loss function of the lth enhancement layer can be expressed as:

[0211]

[0212] In the formula, β l is the connection weight of the lth layer, and λ=0.1 is the regularization parameter. Generally speaking, β l The calculation of can be calculated analytically using the Moore-Penrose pseudo-inverse method or the ridge regression method, expressed as:

[0213]

[0214] In the formula, is the expanded input matrix and I is the identity matrix.

[0215] Step 14: The prediction model implements the dynamic weight integration strategy after the deep network learning is completed. The prediction output of the DWedRVFL network with 10 enhancement layers at time t is The design goal of the DWedRVFL integration module is to find a set of combined prediction weights that minimize the root mean square error (RMSE) between the predicted value and the actual value. The objective function can be expressed as:

[0217]

[0218] Step 15: In order to enhance the adaptability of DWedRVFL to the time-varying characteristics of building loads, DWedRVFL introduces a contribution index based on the latest accuracy, which is defined as a function of the latest measurement prediction error.

[0219]

[0220] In the formula, f l It represents the contribution index of the lth layer based on the latest accuracy at time t-1; taking the third layer as an example, l = 3. Then the calculated f3 = 0.0475. In order to prevent the weight of the dominant output layer from being too large, the contribution index values ​​of each layer are sorted from large to small using the sorting method, and the following order is obtained:

[0221] f1≥f2≥…≥f3≥…≥f L ,

[0222] Subsequently, the latest accuracy contribution of layer 3 is ranked as:

[0223]

[0224] in It represents the index value after ranking at the third layer, and F3=4 represents the ranking of its contribution to the latest accuracy.

[0225] Step 16: DWedRVFL promotes the diversity of overall predictions by giving greater weight to the output layers with diversity contributions. First, the Euclidean distance between the third layer prediction value and the predictions of other layers is defined as the diversity contribution of the third layer:

[0226]

[0227] Among them, E3 = 2.9004 is the sum of the Euclidean distances between the third layer prediction and the predictions of other layers at time t. According to the predicted distribution, E3 is defined as the diversity contribution index. Then, the layers are sorted in descending order of their contribution to diversity, and the resulting order is as follows:

[0228] E1≥E2≥…≥E3≥…≥E L ,

[0229] The third-tier diversity ranking is assigned as follows:

[0230]

[0231] In the formula, is the index value of E3 sorted in descending order, Rank the diversity based on the predicted distribution.

[0232] Step 17: To obtain a more comprehensive diversity measure, DWedRVFL further defines diversity based on prediction performance. The diversity contribution index based on prediction performance is calculated as the difference between a single prediction and the median of all predictions. When calculating the deviation, the true label at time t is unknown. Therefore, we use the latest prediction error to calculate the deviation. Then the deviation of layer 3 at time (t-1) is for:

[0233]

[0234] Will Sort in descending order to get the ranking of layer 3:

[0235]

[0236]

[0237] in The ranking indicators based on prediction performance and diversity ranking are 2 and 7 respectively. The overall diversity contribution ranking of Tier 3 is defined as:

[0238]

[0239] Among them, d3=0.0623 is the overall diversity contribution index of the third layer. The index value of d3 after reverse sorting is equal to 3. D3 is the 3rd layer and its ranking based on comprehensive diversity index is 6.

[0240] Step 18: Based on the ranking information of F3 and D3, the final contribution index r3 of the third output layer is defined as:

[0241] r3=αF3+(1-α)D3,

[0242] Among them, α is a trade-off parameter, and its value is 0.6, so the final contribution index value of the third layer is 4.8.

[0243] Step 19: DWedRVFL uses a ranking-based strategy to calculate the final ranking R3 and combination weights of layer 3 .

[0244] r1≥r2≥…≥r3≥…≥r L ,

[0245]

[0246] in The index value of r3 after descending is equal to 2,

[0247] Step 20: Calculate the combined prediction value at time t based on the sorting strategy According to the calculation method of the third layer output layer weight, the combined weights of all layers are (0.1818 0.0727 0.0364 0.0182 0.0545 0.09090.1091 0.1636 0.1273 0.1455), and the final combined prediction value is calculated as follows:

[0248]

[0249] The final combined prediction value is 1.1567, the true value is 1.9200, and the absolute error is 0.7633. The basic edRVFL uses the median combination of the prediction values ​​of each layer, and the combined prediction value obtained is 1.0969, with an absolute error of 0.8231. Therefore, DWedRVFL is better than the basic edRVFL. The ranking-based dynamic integration strategy it adopts provides a balanced combination weight allocation scheme, thereby generating a more reasonable combined prediction value. The trained DWedRVFL is used for the prediction of the test set, and the prediction results are as follows: Fig.13 As shown:

[0250] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0251] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0252] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0253] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0254] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A building load forecasting method based on forward decomposition denoising and deep ensemble learning, characterized in that: For the original load data, the CEEMDAN algorithm is first used to forward decompose the load data into multiple IMFs components. After removing the high-frequency noise components, the remaining components are reconstructed and fused to obtain clean load data; then, in order to cope with the strong volatility and time-varying nature of building load data, a DWedRVFL neural network model is constructed. The DWedRVFL neural network model uses the latest prediction accuracy to capture the dynamic changes of the load sequence, and constructs a diversity index to utilize the multi-scale features of different output layers; in addition, a sorting strategy is constructed to integrate the contributions of the latest accuracy and diversity, while avoiding the impact of abnormal prediction values ​​on the combined prediction; then, the connection matrix composed of historical load, weather conditions and occupancy data is fed into the DWedRVFL neural network model for learning, thereby improving the accuracy and stability of building load prediction.

2. A building load forecasting method based on forward decomposition denoising and deep ensemble learning according to claim 1, characterized in that: The implementation method of the CEEMDAN algorithm is: Step 1: Assume that the collected original load data is The forward rolling decomposition method is used instead of the decomposition of the entire signal. That is, CEEMDAN decomposes the data within a forward rolling window size w. The data within w only includes known load data and does not involve future data. The load data used for decomposition is expressed as Step 2: Add Gaussian white noise sequence; in the original load data S L Add a set of Gaussian white noise sequences to form new data S new , as shown below: S new =S L +γ0G i Where i is the number of trials; G i is a normally distributed Gaussian noise sequence; γ0 is the ratio of data to noise, which is used to control the additional noise of the original data set; Step 3: According to the EMD method, S new Decompose; from the signal S new Extract the local maximum and minimum points, and calculate the mean M of the upper envelope U and the lower envelope L according to the local extreme points: Step 4: From the original signal S new Subtract the mean M from the modal component h1 to obtain the preliminary modal characteristic component h1: h1=S new -M Step 5: Check whether h1 satisfies the two conditions that the number of zero crossing points should be equal to the number of extreme points and the amplitude is symmetric; if not, it is regarded as a new signal, and steps 3 and 4 are repeated until the IMF conditions are met. After convergence, the first component of EMD is initially obtained; then the first component of CEEMDAN, IMF1, is obtained by calculating the mean, as shown below: Where T represents the total number of times white noise is added; Step 6: Calculate the first residual term r1 as follows: r1=S L -IMF1 Step 7: Add noise to the first residual term r1 to obtain a new sequence r1+γ1E1(G i ), the new sequence is decomposed as follows to obtain the second component IMF2 and the second residual term r2: r2=r1-IMF2 Step 8: Calculate the k-order residual term r based on the same method as step 7 k and the k+1th order component IMF k+1 : r k =r k-1 -IMF k Step 9: Repeat step 8 until the residual term can no longer be decomposed. The final residual term R is calculated as follows: Step 10: Assuming that component IMF1 is the high-frequency noise component that needs to be removed, the remaining components except IMF1 and the residual term R are reconstructed into clean load data: Step 11: X L represents the denoised historical load sequence, X n and X w They represent the building occupancy data obtained from the punch card system and the weather forecast data obtained from the BEMS system respectively; the prediction modeling goal is to establish the following mapping relationship: In the formula, is the predicted value at the next moment, e is the error; the rolling window w determines the amount of historical load data used; among them, the historical load series is expressed as And X n , X W is a scalar; [X L ,X n ,X W ] forms a connection matrix X as the input of the prediction model, and the output is the predicted load value of the next time step.

3. A building load forecasting method based on forward decomposition denoising and deep ensemble learning according to claim 2, characterized in that: The implementation method of the DWedRVFL neural network model is: The input data is Where N and m are the number of samples and features of the input data; P is defined as the number of neurons in the enhancement layer; the feature representation of the first enhancement layer of DWedRVFL is Calculated by the following formula: H 1 =g(XB1) Where g(·) represents a nonlinear activation function, Randomly generated connection weights for the first enhancement layer; For the DWedRVFL model with L enhancement layers, the input of the lth enhancement layer consists of the initial input feature X and the output H of the previous enhancement layer. l-1 The calculation of the lth enhancement layer can be defined as follows: H l =g([H l-1 ,X]B l ) in, is the weight and enhancement feature of the l-th enhancement layer, is the output of the lth layer; then the loss function of the lth enhancement layer can be expressed as: In the formula, β l is the connection weight of the lth layer, λ is the regularization parameter; generally speaking, β l The calculation of can be calculated analytically using the Moore-Penrose pseudo-inverse method or the ridge regression method, expressed as: In the formula, is the extended input matrix, I is the identity matrix; each RVFL module in DWedRVFL corresponds to a decision subtask, and the output weights of each layer are calculated separately; all weights of the hidden layers of DWedRVFL are randomly generated before the start of training and remain fixed during the network training process; In view of the high volatility and time-varying characteristics of building loads, a dynamic integration module is designed. The outputs of different enhancement layers are combined by dynamically updating weights to achieve a trade-off between the latest accuracy and diversity. The predicted output of the network with L enhancement layers is The design goal of the integration module is to find a set of combined weights that minimize the root mean square error (RMSE) between the predicted value and the actual value. The objective function is expressed as:

4. A building load forecasting method based on forward decomposition denoising and deep ensemble learning according to claim 3, characterized in that: In order to enhance the adaptability of DWedRVFL to the time-varying characteristics of building loads, DWedRVFL introduces a contribution index based on the latest accuracy, which is defined as a function of the latest measurement prediction error: In the formula, f l It represents the contribution index of the lth layer based on the latest accuracy at time t-1; in order to prevent the weight of the dominant output layer from being too large, the contribution index values ​​of each layer are sorted from large to small using the sorting method, and the following order is obtained: f1≥f2≥…≥f l ≥…≥f L Then, the latest accuracy contribution of layer l is sorted as: in, represents the index value after sorting at the lth level, F l Indicates the ranking of its contribution to the latest accuracy; if the latest accuracy contribution of the lth layer is the largest, its ranking index value is 1, F l equal to L; for each output layer, the ranking value F l The higher it is, the greater its contribution to the latest accuracy, which means it has a higher level of accuracy.

5. A building load forecasting method based on forward decomposition denoising and deep ensemble learning according to claim 4, characterized in that: DWedRVFL introduces diversity to ensure the effective utilization of multi-scale enhanced features from different output layers, thereby effectively coping with the volatility of building loads; DWedRVFL promotes the diversity of overall predictions by giving greater weight to the output layers with diversity contributions; first, the Euclidean distance between the predictions of each layer and the predictions of other layers is defined as the diversity contribution: Among them, E l is the sum of the Euclidean distances between the prediction of the lth layer and the predictions of other layers at time t. According to the prediction distribution, E l Defined as the diversity contribution index; then sort the diversity contribution index in descending order, and the order obtained is as follows: E1≥E2≥…≥E l ≥…≥E L Then the diversity ranking of the lth layer is assigned as: In the formula, For E l The index after sorting in descending order, is the diversity ranking based on the predicted distribution; for the output layer l, the greater its diversity contribution, The bigger; In order to obtain a more comprehensive diversity metric, DWedRVFL further defines diversity based on prediction performance. The diversity contribution index based on prediction performance is calculated as the difference between a single prediction and the median of all predictions. When calculating the deviation, the true label at time t is unknown. Therefore, the latest prediction error is used to calculate the deviation. The deviation of the lth layer at time (t-1) is for: Will Sort in descending order to get the ranking of the lth layer: in, and are the index based on prediction performance and the diversity contribution ranking respectively; according to the above two diversity calculation results, the overall diversity contribution ranking of the lth layer is defined as: d1≥d2≥…≥d l ≥…≥d L Among them, d l is the overall diversity contribution indicator for layer l, is d l Index value after reverse sorting; D l is the ranking value of the lth layer based on the comprehensive diversity index; Comprehensive F l and D l The sorting information defines the final contribution index r of the lth output layer l for: r l =αF l +(1-a)D l Among them, α is the trade-off parameter; DWedRVFL uses a ranking-based strategy to calculate the final ranking R of the lth layer l and combination weight r1≥r2≥…≥r l ≥…≥r L , in, For r l Index value after descending order; Calculate the combined prediction value of DWedRVFL at time t according to the sorting strategy The calculation method is:

Images