Building energy consumption prediction method based on entropy evaluation and related device

Abstract

The invention discloses a building energy consumption prediction method based on entropy evaluation and a related device, and the method comprises the steps: obtaining building data; decomposing the building data to obtain a plurality of subsequences; calculating the permutation entropy value of each sub-sequence; determining a prediction model corresponding to each sub-sequence based on the permutation entropy value of each sub-sequence; and obtaining the building energy consumption based on each sub-sequence and the prediction model corresponding to each sub-sequence. According to the method, a single prediction method is abandoned, the prediction model corresponding to each sub-sequence is determined based on the permutation entropy, and the adaptation degree between the prediction model and each sub-sequence is improved, so that the influence of non-stationary information of each sub-sequence on prediction is effectively isolated, and the prediction accuracy is improved. And the accuracy of obtaining the building energy consumption based on each sub-sequence and the prediction model corresponding to each sub-sequence is improved.

Description

Technical Field

[0001] The present application relates to the technical field of energy management, and particularly to a building energy consumption prediction method and related devices based on entropy evaluation. Background Art

[0002] With the rapid development of the construction industry, building energy planning and management have gradually become the focus of attention. Building energy consumption prediction plays a crucial role in building energy planning and management. Building energy consumption prediction is the basis for realizing advanced building energy management technologies such as fault detection, demand response, and optimal control. Energy consumption prediction is divided into four categories: long-term, medium-term, short-term, and very short-term. Short-term and very short-term predictions can effectively alleviate the contradiction between energy supply and demand and improve the economy, stability, and security of the energy system. Medium-term and long-term predictions can provide strong decision-making support for long-term energy planning and deployment. Therefore, accurate prediction of building energy consumption is crucial for the development of the construction industry.

[0003] Although many effective models have been introduced in the current field of building energy consumption prediction, building energy consumption data often exhibits non-stationarity and high volatility. A single prediction method cannot effectively isolate the impact of these non-stationary information on the overall prediction, resulting in low accuracy of building energy consumption prediction. Summary of the Invention

[0004] Based on the above problems, the present application provides a building energy consumption prediction method and related devices based on entropy evaluation, aiming to improve the accuracy of building energy consumption prediction.

[0005] The embodiments of the present application disclose the following technical solutions:

[0006] The first aspect of the present application provides a building energy consumption prediction method based on entropy evaluation, including:

[0007] Obtain building data;

[0008] Decompose the building data to obtain multiple subsequences;

[0009] Calculate the permutation entropy value of each subsequence;

[0010] Determine the prediction model corresponding to each subsequence based on the permutation entropy value of each subsequence; the prediction model is an artificial neural network deep learning prediction model, a long short-term memory artificial neural network deep learning prediction model, a convolutional neural network-long short-term memory artificial neural network deep learning prediction model, or a temporal convolutional network model;

[0011] Obtain the building energy consumption based on each subsequence and the prediction model corresponding to each subsequence.

[0012] Optionally, decomposing the building data to obtain multiple subsequences specifically includes:

[0013] Determine the wavelet basis function and the decomposition threshold;

[0014] Perform three - level decomposition processing on the building data based on the wavelet basis function and the decomposition threshold to obtain three subsequences.

[0015] Optionally, performing three - level decomposition processing on the building data based on the wavelet basis function and the decomposition threshold to obtain three subsequences specifically includes:

[0016] Let the value of k be 1; set the building data as the building sequence of the first iteration;

[0017] Decompose the building sequence of the k - th iteration through the wavelet basis function to obtain a set of components; the set of components includes a first component, a second component, a third component, and a fourth component; the frequency of the first component is lower than that of the second component; the frequency of the second component is lower than that of the third component; the frequency of the third component is lower than that of the fourth component;

[0018] Set the coefficients in the fourth component to 0, and perform soft - thresholding processing on the coefficients in the first component, the second component, and the third component respectively based on the decomposition threshold to obtain a processed first component, a processed second component, a processed third component, and a processed fourth component;

[0019] Perform inverse wavelet transform on the processed first component, the processed second component, the processed third component, and the processed fourth component to obtain the subsequence of the k - th iteration, and store the subsequence of the k - th iteration in the subsequence set;

[0020] Judge whether the iteration end condition is satisfied to obtain a first judgment result; the iteration end condition is that k is equal to 3;

[0021] If the first judgment result is yes, end the iteration and output the subsequences in the subsequence set;

[0022] If the first judgment result is no, subtract the building sequence of the k - th iteration from the subsequence of the k - th iteration to obtain the residual sequence of the k - th iteration;

[0023] Set the residual sequence of the k - th iteration as the building sequence of the (k + 1)-th iteration, let the value of k increase by 1, and return to the step of decomposing the building sequence of the k - th iteration through the wavelet basis function to obtain a set of components.

[0024] Optionally, determining a prediction model corresponding to each subsequence based on the permutation entropy value of each subsequence specifically includes:

[0025] Determining a prediction model corresponding to each subsequence based on a preset condition and the permutation entropy value of each subsequence; the preset condition includes using an artificial neural network model as the prediction model for the subsequence corresponding to the permutation entropy value satisfying the first preset range, using a long short-term memory network model as the prediction model for the subsequence corresponding to the permutation entropy value satisfying the second preset range, using a CNN-LSTM neural network model as the prediction model for the subsequence corresponding to the permutation entropy value satisfying the third preset range, and using a temporal convolutional network model as the prediction model for the subsequence corresponding to the permutation entropy value satisfying the fourth preset range.

[0026] Optionally, obtaining building energy consumption based on each subsequence and the prediction model corresponding to each subsequence specifically includes:

[0027] Inputting each subsequence into the corresponding prediction model to obtain a prediction result corresponding to each subsequence;

[0028] Weighted summing the prediction results corresponding to each subsequence to obtain the predicted building energy consumption.

[0029] Optionally, the obtaining of building data specifically includes:

[0030] Obtaining building data to be processed;

[0031] Judging whether there are outliers in the building data to be processed to obtain a second judgment result;

[0032] If the second judgment result is yes, using the seasonal hybrid ESD method to replace the outliers with missing values, calculating the mean of the data before and after the missing values, and using the mean to replace the missing values to obtain the building data;

[0033] If the second judgment result is no, using the building data to be processed as the building data.

[0034] The second aspect of the present application provides a building energy consumption prediction device based on entropy evaluation, including:

[0035] An obtaining module, configured to obtain building data;

[0036] A decomposition module, configured to decompose the building data to obtain multiple subsequences;

[0037] A permutation entropy value calculation module, configured to calculate the permutation entropy value of each subsequence;

[0038] A prediction model determination module, configured to determine a prediction model corresponding to each subsequence based on the permutation entropy value of each subsequence; the prediction model is an artificial neural network deep learning prediction model, a long short-term memory artificial neural network deep learning prediction model, a convolutional neural network-long short-term memory artificial neural network deep learning prediction model, or a temporal convolutional network model;

[0039] A building energy consumption determination module, configured to obtain the building energy consumption based on each subsequence and the prediction model corresponding to each subsequence.

[0040] A third aspect of the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the building energy consumption prediction method based on entropy evaluation provided in the first aspect.

[0041] A fourth aspect of the present application is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the building energy consumption prediction method based on entropy evaluation provided in the first aspect.

[0042] A fifth aspect of the present application is a computer program product, including a computer program, and when the computer program is executed by a processor, it implements the building energy consumption prediction method based on entropy evaluation provided in the first aspect.

[0043] Compared with the prior art, the present application has the following beneficial effects:

[0044] The present application decomposes building data to obtain multiple subsequences; calculates the permutation entropy value of each subsequence; and determines a prediction model corresponding to each subsequence based on the permutation entropy value of each subsequence. The present application abandons a single prediction method, and by determining a prediction model corresponding to each subsequence based on the permutation entropy value, the adaptability between the prediction model and each subsequence is improved, thereby effectively isolating the influence of the non-stationary information of each subsequence on the prediction, and further improving the accuracy of obtaining the building energy consumption based on each subsequence and the prediction model corresponding to each subsequence. Description of the Drawings

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0046] Figure 1 It is a flowchart of a building energy consumption prediction method based on entropy evaluation provided in an embodiment of the present application;

[0047] Figure 2 This is a flowchart of the wavelet threshold decomposition method provided by the embodiments of the present application;

[0048] Figure 3 This is a schematic diagram of the network model structure of the artificial neural network deep learning prediction model provided by the embodiments of the present application;

[0049] Figure 4 This is a schematic diagram of the network model structure of the long short-term memory artificial neural network deep learning prediction model provided by the embodiments of the present application;

[0050] Figure 5 This is a schematic diagram of the network model structure of the convolutional neural network-long short-term memory artificial neural network deep learning prediction model provided by the embodiments of the present application;

[0051] Figure 6 This is a schematic diagram of the causal neural network structure of the TCN with dilated convolutions provided by the embodiments of the present application;

[0052] Figure 7 This is a schematic diagram of the residual module structure of the TCN deep learning prediction model provided by the embodiments of the present application;

[0053] Figure 8 This is a schematic diagram of the determination of the prediction model provided by the embodiments of the present application;

[0054] Figure 9 This is a structural diagram of a building energy consumption prediction device based on entropy evaluation provided by the embodiments of the present application. Detailed implementation manners

[0055] As described above, building energy consumption data often exhibits instability and high volatility. Using a single prediction method cannot effectively isolate the impact of instability on prediction, resulting in relatively low accuracy of building energy consumption prediction.

[0056] The inventors have found through research that decomposing building data to obtain multiple subsequences; calculating the permutation entropy value of each subsequence; determining the prediction model corresponding to each subsequence based on the permutation entropy value, improving the fitness between the prediction model and each subsequence, thereby effectively isolating the impact of the non-stationary information of each subsequence on prediction, and further improving the accuracy of building energy consumption obtained based on each subsequence and the prediction model corresponding to each subsequence.

[0057] To enable those skilled in the art to better understand the solution of this application, the following will clearly and completely describe the technical solution in the embodiments of this application in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts shall fall within the scope of protection of this application.

[0058] Figure 1 The flowchart of a building energy consumption prediction method based on entropy evaluation provided by the embodiments of this application is now combined with Figure 1 To illustrate a building energy consumption prediction method based on entropy evaluation:

[0059] S101: Obtain building data.

[0060] Here, the building data may include one or more of various types of building data such as building name, building area, building materials, average energy consumption, etc.

[0061] Due to emergencies such as equipment failures, extreme climates, or data transmission interruptions, problems such as abnormal or missing building data may occur, and appropriate preprocessing measures need to be taken. This application provides an optional embodiment:

[0062] Obtain the building data to be processed.

[0063] The building data to be processed represents building data that may have problems such as abnormal data or missing data.

[0064] Judge whether there are outliers in the building data to be processed to obtain a second judgment result.

[0065] If the second judgment result is yes, use the seasonal hybrid ESD method to replace the outliers with missing values, calculate the mean of the data before and after the missing values, and use the mean to replace the missing values to obtain the building data.

[0066] Using the mean of the data before and after the missing values to replace the missing values can ensure the accuracy of the replacement of the missing values and avoid a large gap between the supplemented data and the original data.

[0067] If the second judgment result is no, use the building data to be processed as the building data.

[0068] After obtaining the building data, it is necessary to predict the building energy consumption based on the building data. However, since the building data often exhibits instability and high volatility, it is necessary to separate the building data to effectively isolate the unstable information in the building data and improve the prediction accuracy. The existing technology separates the building data by using traditional data decomposition methods, and then uses a single prediction method to predict the separated sequences. The traditional data decomposition methods include Empirical Mode Decomposition (EMD), Ensemble Empirical Mode Decomposition (EEMD), and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN).

[0069] EMD can decompose the building data into multiple Intrinsic Mode Functions (IMFs) and a residual component. Among them, each decomposed IMF only exhibits a single instantaneous frequency and must satisfy two constraints: the mean of the upper and lower envelope lines is 0, and the difference between the number of extreme points and zero points is less than or equal to 1. EMD often has a relatively obvious mode mixing problem, resulting in an unsatisfactory final decomposition effect.

[0070] EEMD effectively solves the mode mixing problem by adding white noise to automatically distribute the signal to an appropriate reference scale. However, there is a certain residue of white noise in the decomposed IMFs, which affects the reconstruction effect.

[0071] Based on EEMD, CEEMDAN introduces limited adaptive white noise and cancels the noise through multiple superpositions and averages, solving the mode mixing problem faced by EMD. And compared with EEMD, CEEMDAN has less residual noise in the reconstructed signal and higher decomposition efficiency. Therefore, CEEMDAN is mostly used in the existing technology to separate the building data and isolate the non-stationary information in the building data. However, CEEMDAN still has some problems. Now, the existing technology combining CEEMDAN with a single prediction method will be described in detail:

[0072] Using CEEMDAN to decompose the building data into multiple subsequences, the specific steps are as follows:

[0073] Add Gaussian white noise n i (t) to the building data X(t):

[0074] X i (t) = X(t) + ω 0 ni (t)

[0075] where ω 0 is the noise figure.

[0076] Let the total number of times of adding white noise be n, and all the data X i (t) (i = 1, 2, …, n) are used for the first decomposition to obtain the corresponding IMF 1 (X i (t)), and the average value is obtained to get the intrinsic mode function IMF 1 . The corresponding formula is:

[0077]

[0078] Subtract X(t) from the IMF 1 component to obtain the first residual sequence r 1 (t):

[0079] r 1 (t) = X(t) - IMF 1

[0080] Then continue to add white noise to r 1 (t) to obtain the second intrinsic mode function IMF 2 and the second residual sequence r 2 (t):

[0081]

[0082] r 2 (t) = r 1 (t) - IMF 2

[0083] where EMD j () represents the jth IMF component obtained from the EMD decomposition.

[0084] Repeat the following process:

[0085]

[0086] r k' (t) = r k'-1 (t) - IMF k '

[0087] Until the residual component obtained at the k'th decomposition is a monotonic function and cannot be decomposed, the CEEMDAN decomposition terminates. At this time, the k' IMF components obtained by the decomposition and X(t) satisfy the following relationship:

[0088]

[0089] The k IMF components obtained by decomposition are used for building energy consumption prediction by a single prediction method.

[0090] There are two problems in the existing technology. First, predicting multiple decomposed subsequences will incur a large amount of time overhead, which will limit the application in scenarios with high timeliness requirements and CEEMDAN has the problem of over-decomposition. Second, since the decomposition algorithm will generate many subsequences with different characteristics, it is difficult for a single prediction model to adapt to all subsequences. Both of the two problems existing in the existing technology will lead to low accuracy of building energy consumption prediction.

[0091] Therefore, in order to improve the accuracy of building energy consumption prediction, it is necessary to improve the existing technology from two aspects. The first aspect is to optimize the traditional data decomposition method, and the second aspect is to allocate appropriate prediction models for each subsequence.

[0092] In the first aspect, it is necessary to conduct a systematic analysis of the traditional data decomposition method, summarize the necessary factors and non-necessary factors for the traditional data decomposition method to improve the prediction accuracy, and design an optimization scheme for the decomposition algorithm based on this.

[0093] This application provides an analysis example. The historical building data of a building is selected from the public dataset as the original data, and CEEMDAN is used to decompose the original data, and a total of 13 subsequences IMF1-IMF13 can be obtained. Calculate the normalized permutation entropy value of each subsequence. Calculate the average value of the absolute value of each subsequence, which is called the amplitude mean of this sequence. Calculate the average value of the original data as the amplitude mean of the original data. Then calculate the proportion of the amplitude mean of each subsequence to the amplitude mean of the original sequence.

[0094] Analysis and comparison show that after the original sequence is decomposed by CEEMDAN, the most important trend information is concentrated in the simple sequences after decomposition. These sequences hardly have non-stationary information. Therefore, the existing prediction models can easily achieve satisfactory results for the prediction of such sequences. Most of the non-stationary information in the original sequence exists in the complex sequences and chaotic sequences obtained after decomposition, but their proportion of the amplitude mean is very small. These sequences only fluctuate within a very small range and have little impact on the overall prediction result.

[0095] Still taking CEEMDAN as an example, the optimization scheme of the decomposition algorithm is further discussed. The IMFs from IMF3 to IMF13 in the decomposed subsequences are reconstructed. The reconstruction method destroys the constraint conditions of the intrinsic mode function by adding the subsequences, and also causes aliasing of the subsequences of different modes. Finally, only three subsequences, namely IMF1, IMF2, and the reconstructed sequence, are retained and renamed as subsequence 1, subsequence 2, and subsequence 3. Predict the three subsequences after decomposition and reconstruction, and analyze and compare the prediction effects of the three subsequences. It is found that the reconstructed prediction method shows higher accuracy and efficiency than the decomposition method without reconstruction. Therefore, the decomposition algorithm should pay more attention to how to separate non-stationary information more comprehensively, rather than over-decomposing simple sequences that are already easy to predict and increasing the training burden.

[0096] Therefore, by observing the frequency and amplitude characteristics of the subsequences after decomposition and reconstruction, it is concluded that the main reason for effectively improving the prediction accuracy of the decomposition algorithm is to separate non-stationary information from the original sequence and shrink it to a relatively small range, weakening its impact on the overall prediction. The process of decomposition and reconstruction will not reduce the prediction accuracy of the hybrid prediction method. In response to this series of problems, this application designs a more efficient wavelet threshold decomposition method (Wavelet Thresholding Decomposition, WTD). The original data is converted into frequency-domain information, and after processing the frequency-domain information through low-pass filtering and thresholding, it is then converted back into time-domain information to generate three subsequences, achieving a result similar to the original decomposition and reconstruction process, but greatly shortening the decomposition time and prediction time, and improving the overall prediction efficiency. Now, a detailed description of WTD is as follows:

[0097] S102: Decompose the building data to obtain multiple subsequences.

[0098] As an alternative embodiment, Figure 2 is the flowchart of the wavelet threshold decomposition method provided by the embodiments of this application. As Figure 2 shown, determine the wavelet basis function and the decomposition threshold.

[0099] This application also provides a specific formula to solve the decomposition threshold T:

[0100]

[0101] Among them, the threshold T reflects the fluctuation of the data in terms of the average amplitude. Based on the wavelet basis function and the decomposition threshold, the building data is decomposed three times to obtain three subsequences. The three subsequences represent the trend component, the detail component, and the residual component respectively.

[0102] This application provides an alternative embodiment of decomposing the building data three times:

[0103] Set the value of k to 1; set the building data as the building sequence of the first iteration.

[0104] Decompose the building sequence of the k-th iteration through wavelet basis functions to obtain a set of components; the set of components includes a first component, a second component, a third component, and a fourth component; the frequency of the first component is lower than that of the second component; the frequency of the second component is lower than that of the third component; the frequency of the third component is lower than that of the fourth component.

[0105] As an alternative embodiment, decompose the building sequence three times through the convolution and downsampling operations of wavelet basis functions. Decompose the building sequence (time series) for the first time to obtain cA 1 and cD 1 , decompose cA 1 for the second time to obtain cA 2 and cD 2 , decompose cA 2 for the third time to obtain cA 3 and cD 3 . After three decompositions, obtain four components with increasing frequencies, namely the first component cA 3 , the second component cD 3 , the third component cD 2 , and the fourth component cD 1 .

[0106] Set the coefficients in the fourth component cD 1 to 0, and perform soft thresholding on the coefficients in the first component cA 3 , the second component cD 3 , and the third component cD 2 respectively based on the decomposition threshold to obtain the processed first component, the processed second component, the processed third component, and the processed fourth component.

[0107] As an alternative embodiment, the following formula can be used to perform soft thresholding on each coefficient W in cA 3 , cD 3 , cD 2 to obtain W t :

[0108]

[0109] Perform inverse wavelet transform on the processed first component, the processed second component, the processed third component, and the processed fourth component to obtain the subsequence of the k-th iteration, and store the subsequence of the k-th iteration in the subsequence set.

[0110] Judge whether the iteration end condition is satisfied to obtain a first judgment result; the iteration end condition is that k is equal to 3.

[0111] If the first judgment result is yes, the iteration ends and the subsequence in the subsequence set is output.

[0112] If the first judgment result is no, the building sequence of the kth iteration is subtracted from the subsequence of the kth iteration to obtain the residual sequence of the kth iteration.

[0113] The residual sequence of the kth iteration is set as the building sequence of the k+1th iteration, the value of k is increased by 1, and the step of decomposing the building sequence of the kth iteration by the wavelet basis function to obtain the component set is returned.

[0114] like Figure 2 As shown, WTD is obtained by performing low-pass filtering and thresholding on the time series to obtain a subsequence and residual representing the trend component, and by performing low-pass filtering and thresholding on the residual to obtain a subsequence representing the detail component and a subsequence representing the residual component.

[0115] To prove the effectiveness of WTD, this application combines WTD with an artificial neural network deep learning prediction model, a long short-term memory artificial neural network deep learning prediction model, a convolutional neural network-long short-term memory artificial neural network deep learning prediction model and a temporal convolutional network model to design a new hybrid prediction method. According to the different types of prediction models, they are WTD-ANN, WTDLSTM, WTD-CNN-LSTM and WTD-TCN, respectively. It is also comprehensively compared in terms of accuracy and efficiency with the hybrid prediction methods based on the combination of CEEMDAN with an artificial neural network deep learning prediction model, a long short-term memory artificial neural network deep learning prediction model, a convolutional neural network-long short-term memory artificial neural network deep learning prediction model and a temporal convolutional network model (CEEMDAN-ANN, CEEMDAN-LSTM, CEEMDAN-CNN-LSTM, CEEMDAN-TCN). It can be found that, unlike the decomposition time of hundreds of seconds of CEEMDAN, the decomposition time of WTD can be shortened to milliseconds. It can be seen that the use of WTD can greatly improve the prediction efficiency of the hybrid prediction method. At the same time, no matter which prediction model is used, the hybrid prediction method using WTD is far ahead of the hybrid prediction method of the same level using CEEMDAN in terms of prediction accuracy, which fully demonstrates the effectiveness and advancement of WTD.

[0116] As an optional embodiment, Figure 3 A schematic diagram of the network model structure of the artificial neural network deep learning prediction model provided in the embodiment of the present application, such as Figure 3 As shown in Figure 1, the network model structure of the artificial neural network deep learning prediction model includes an input layer, a hidden layer, and an output layer. The input layer converts X 0 , X 1, X 2 to X t are respectively transmitted to the hidden layer, and the hidden layer processes X 0 , X 1 , X 2 to X t to obtain the final prediction result, and the prediction result is transmitted to the output layer, and the output layer outputs the prediction result y t .

[0117] As an alternative embodiment Figure 4 is a schematic diagram of the network model structure of the long short-term memory artificial neural network deep learning prediction model provided by the embodiment of the present application, as Figure 4 shown, the network model structure of the long short-term memory artificial neural network deep learning prediction model includes 3 sigmoids (forget gate, input gate, and output gate) and two tanhs. The data processing flow of the long short-term memory artificial neural network deep learning prediction model is: input X t into the long short-term memory artificial neural network deep learning prediction model, and the forget gate, input gate, and output gate respectively obtain the forget gate output, input gate output, and output gate output based on the hidden state h t-1 at the previous time step and X t ; process h t-1 and X t using tanh to obtain the candidate value of the new information; obtain the updated cell state C t-1 based on the forget gate output, the cell state C t at the previous time step, the input gate output, and the candidate value of the new information; use tanh to obtain h t based on the output gate output and C t .

[0118] As an alternative embodiment Figure 5 is a schematic diagram of the network model structure of the convolutional neural network-long short-term memory artificial neural network deep learning prediction model provided by the embodiment of the present application, as Figure 5 shown, the convolutional neural network-long short-term memory artificial neural network deep learning prediction model includes a convolutional layer, an LSTM layer, and a fully connected layer; the data processing flow of the convolutional neural network-long short-term memory artificial neural network deep learning prediction model is: input data into the convolutional layer, the convolutional layer processes the data, and transmits the processed data to the LSTM layer, and the LSTM layer processes the data and then uses the fully connected layer for connection to obtain the output data

[0119] As an alternative embodiment Figure 6 is a schematic diagram of the causal neural network structure of the TCN with dilated convolutions provided by the embodiment of the present application, as Figure 6As shown, the data processing flow of the causal neural network of TCN with dilated convolution is as follows: the input layer inputs the sequence X 0 、X 1 、X 2 to X t ; the hidden layer d = 1 uses a convolutional kernel with a dilation factor of 1 to perform a convolutional operation on the sequence and transmits it to the hidden layer d = 2; the hidden layer d = 2 uses a convolutional kernel with a dilation factor of 2 to perform a convolutional operation on the sequence and transmits it to the hidden layer d = 4; the hidden layer d = 4 uses a convolutional kernel with a dilation factor of 4 to perform a convolutional operation on the sequence and transmits it to the output layer; the output layer outputs the sequence y 0 、y 1 、y 2 to y t .

[0120] As an alternative embodiment, Figure 7 is a schematic diagram of the residual module structure of the TCN deep learning prediction model provided by the embodiment of the present application. The data processing flow of the residual module of the TCN deep learning prediction model is as follows: the input sequence The dilated causal convolution performs a causal convolution operation using the dilation factor; the weight normalization normalizes the output after convolution; the ReLU activation function applies the ReLU activation function to activate the normalized data; Dropout applies Dropout to process the activated data to prevent overfitting; the dilated causal convolution performs a causal convolution operation again using the dilation factor; the weight normalization normalizes the output after convolution; the ReLU activation function applies the ReLU activation function to activate the normalized data; Dropout applies Dropout again to process the activated data to prevent overfitting. Add to the processed data to obtain the output sequence If the dimensions of the input sequence and the output sequence are inconsistent, a 1×1 convolutional layer can be used to adjust the dimensions.

[0121] After determining the WTD, it is necessary to allocate a suitable prediction model for each subsequence. In this application, the prediction model corresponding to each subsequence is mainly determined by the permutation entropy value. Because in analyzing the influence of sequence entropy value on model prediction, by comprehensively comparing the performance differences shown by different models as the sequence features change, it can be seen that selecting the prediction model with the best comprehensive performance for subsequences with different features is of great significance for improving the overall prediction accuracy and efficiency of the hybrid prediction method.

[0122] Regarding the complexity of the sequence as its main feature and introducing the permutation entropy value as a quantitative index for sequence complexity is because, compared with other entropy value measurement methods, the permutation entropy value has the advantages of simple calculation, fast speed, strong anti-noise ability, etc., and has a wide range of applications in the fields of fault detection, financial analysis, biomedicine, etc. Therefore, in this application, a prediction model is determined through the permutation entropy value, such as calculating the permutation entropy value of each subsequence in S103 and determining the prediction model corresponding to each subsequence based on the permutation entropy value of each subsequence in S104. Now, the calculation process of the permutation entropy value and the determination method of the prediction model corresponding to each subsequence will be described in detail:

[0123] S103: Calculate the permutation entropy value of each subsequence.

[0124] S104: Determine the prediction model corresponding to each subsequence based on the permutation entropy value of each subsequence.

[0125] Regarding the complexity of the sequence as its main feature, introducing the permutation entropy value as a quantitative index for complexity, and calculating the permutation entropy value of each subsequence. This application provides the specific calculation process of the permutation entropy value:

[0126] For a time series X, it can be expressed as:

[0127]

[0128] where T' represents the length of the time series.

[0129] Select appropriate embedding dimension m and time delay τ, perform spatial reconstruction on the time series to obtain a reconstructed matrix of m×l, where l = T'-(m - 1).

[0130] For each vector obtained after reconstruction, it can be expressed as:

[0131] X i =[x(i),x(i + τ),…,x(i + τ(m - 1))] T'

[0132] where 1≤i≤T'-(m - 1)τ. Then, sort each vector according to the magnitude relationship:

[0133] x(i+(j 1 -1)τ)≤x(i+(j 2 -1)τ)≤…≤x(i+(j m -1)τ)

[0134] where 1≤j 1 , j 2 ,…, j m ≤m. If two different elements in the reconstructed vector have the same value, that is:

[0135] x(i+(j a -1)τ) = x(i+(j b -1)τ)

[0136] where i a ≠ i b , sort according to the relative magnitude of i a and i b , from which the sorting index π corresponding to each vector can be obtained i . For example, for the reconstructed vector X i = {3, 6, 3, 2}, its corresponding sorting index is π i = {4, 1, 3, 2}. According to the knowledge of permutations and combinations, at most m! kinds of permutation index combinations can appear. Record that the actual sorting index of the sequence appears J kinds, and define the number of times the j-th sorting appears as v(j). Then the probability p j of the j-th sorting index appearing is:

[0137]

[0138] where N-(m - 1)τ is the number of reconstructed vectors, and it also represents the total number of actual distributions. Finally, in the form of Shannon entropy, calculate the permutation entropy value of this sequence:

[0139]

[0140] After normalization, H PE (m) can be further shrunk to the range of 0 - 1:

[0141]

[0142] When H PE (m) is closer to 1, the sequence is more complex; when it is closer to 0, the sequence is simpler.

[0143] As an alternative embodiment, in order to accelerate the convergence speed of the model and improve the stability of the prediction results, after S103, it can further include using the min-max normalization method to perform planning processing on the permutation entropy value of each subsequence. The corresponding conversion formula is:

[0144]

[0145] where X min and X max are respectively the minimum and maximum values of the energy consumption data, and X norm is the result after normalizing the historical data X.

[0146] After determining the permutation entropy value of each subsequence, a prediction model corresponding to each subsequence is determined based on a preset condition and the permutation entropy value of each subsequence. This application mainly determines the prediction model corresponding to each subsequence through a preset condition, so the preset condition needs to be defined. As an alternative embodiment, the preset condition can be determined through specific experiments; the specific experiment is as follows: Four common deep learning prediction models, ANN, LSTM, CNN-LSTM, and TCN, are used to predict 140 sequence samples in dataset A respectively, and the correlation coefficient R between each prediction result and the true value is calculated. 2 The average value, maximum value, minimum value, and standard deviation of R are statistically calculated according to the entropy value interval where the sequence is located. 2 The prediction accuracy of the model under different entropy value conditions is compared to determine the common law shown by the four prediction models as the sequence entropy value changes, and the sequences are divided into three categories: simple sequences, complex sequences, and chaotic sequences according to the permutation entropy value. By observing the unique performance of each prediction model under different entropy value sequences, the unique advantages and disadvantages of different models in dealing with each entropy value sequence are analyzed in detail, and the preset condition is determined according to the analysis conclusion.

[0147] As an alternative embodiment, the preset condition is set according to the accuracy of the model. The preset condition includes using the artificial neural network model as the prediction model for the subsequence corresponding to the permutation entropy value that satisfies the first preset range, using the long short-term memory network model as the prediction model for the subsequence corresponding to the permutation entropy value that satisfies the second preset range, using the CNN-LSTM neural network model as the prediction model for the subsequence corresponding to the permutation entropy value that satisfies the third preset range, and using the temporal convolutional network model as the prediction model for the subsequence corresponding to the permutation entropy value that satisfies the fourth preset range.

[0148] The specific values of the first preset range, the second preset range, the third preset range, and the fourth preset range can be set according to the analysis conclusion. This application provides an alternative example. Figure 8 This is a schematic diagram for determining the prediction model provided by the embodiment of this application. As shown in Figure 8As shown in the figure, after calculating the normalized permutation entropy value, if the permutation entropy value (PE) is lower than 0.3, the ANN is selected as the prediction model because the accuracy performance of the ANN on low-entropy sequences is not inferior to that of other prediction models. Compared with other prediction models, the ANN also has the advantages of simple structure, low training difficulty, and high prediction efficiency. If the sequence entropy value is between 0.3 and 0.9, the CNN-LSTM is selected as the prediction model because in the comparison of the performance of prediction models for medium-entropy sequences, the CNN-LSTM shows far higher prediction accuracy and stability than other prediction models. If the sequence entropy value is greater than 0.9, the TCN is selected as the prediction model because such sequences contain less effective information and are extremely difficult to predict. Compared with other models, the TCN has more advantages in terms of prediction accuracy and time due to its parallel computing characteristics and strong feature extraction ability.

[0149] This application also provides the correlation coefficient R 2 , and the specific calculation formulas for the root mean square error RMSE, mean absolute error MAE, and mean absolute percentage error MAPE:

[0150]

[0151] Among them, n, y i , respectively represent the true value, predicted value, and the number of true values. The value range of R 2 is from 0 to 1. The closer the value is to 1, the higher the similarity between the predicted sequence and the original data sequence, and the better the model prediction effect. MAE, MAPE, and RMSE show the deviation between the predicted result and the true result in different dimensions. The smaller their values, the better the prediction effect.

[0152] S105: Obtain the building energy consumption based on each subsequence and the prediction model corresponding to each subsequence.

[0153] Input each subsequence into the prediction model corresponding to each subsequence to obtain the prediction result corresponding to each subsequence.

[0154] Sum the prediction results corresponding to each subsequence weighted to obtain the predicted building energy consumption.

[0155] The problem that a single prediction model cannot adapt to the characteristics of most subsequences exists. Therefore, in this application, the unique advantages of each prediction model in processing sequences it is good at are utilized to allocate corresponding prediction models to each subsequence, making the prediction result of building energy consumption much higher in accuracy than the method using a single prediction model. Moreover, this application systematically analyzes the reasons for the effectiveness of the decomposition method, and further designs a wavelet threshold decomposition method through the observation and reconstruction of subsequences. This method greatly improves the decomposition speed and prediction speed. The hybrid prediction method combined with the wavelet threshold decomposition method shows higher prediction accuracy and prediction efficiency compared with the hybrid prediction method combined with other decomposition methods.

[0156] Based on the foregoing embodiments provided, Figure 9 The structure diagram of a building energy consumption prediction device based on entropy evaluation provided by an embodiment of this application is as Figure 9 shown. Correspondingly, this application also provides a building energy consumption prediction device based on entropy evaluation, including:

[0157] An acquisition module, configured to acquire building data.

[0158] The acquisition module includes an acquisition unit, an anomaly judgment unit, and a judgment result unit. Now, the functions of each unit in the acquisition module are described:

[0159] The acquisition unit is configured to acquire building data to be processed.

[0160] The anomaly judgment unit is configured to judge whether there are outliers in the building data to be processed, and obtain a second judgment result.

[0161] The judgment result unit is configured to, if the second judgment result is yes, use the seasonal hybrid ESD method to replace the outliers with missing values, calculate the mean values of the data before and after the missing values, and use the mean values to replace the missing values to obtain the building data. If the second judgment result is no, use the building data to be processed as the building data.

[0162] A decomposition module, configured to decompose the building data to obtain multiple subsequences.

[0163] The decomposition module includes a basis function determination unit and a decomposition unit. Now, the functions of each unit in the decomposition module are described:

[0164] The determination unit is configured to determine the wavelet basis function and the decomposition threshold.

[0165] The decomposition unit is configured to perform three decomposition processes on the building data based on the wavelet basis function and the decomposition threshold to obtain three subsequences.

[0166] The decomposition unit includes a setting subunit, a decomposition subunit, a soft threshold processing subunit, an inverse wavelet transform subunit, a judgment subunit, an output subunit, and an iteration subunit. Now, the functions of each subunit in the decomposition unit will be described:

[0167] The setting subunit is used to set the value of k to 1 and set the building data as the building sequence for the first iteration.

[0168] The decomposition subunit is used to decompose the building sequence for the k-th iteration through a wavelet basis function to obtain a set of components. The set of components includes a first component, a second component, a third component, and a fourth component. The frequency of the first component is lower than that of the second component. The frequency of the second component is lower than that of the third component. The frequency of the third component is lower than that of the fourth component.

[0169] The soft threshold processing subunit is used to set the coefficients in the fourth component to 0 and perform soft thresholding on the coefficients in the first component, the second component, and the third component respectively based on a decomposition threshold to obtain a processed first component, a processed second component, a processed third component, and a processed fourth component.

[0170] The inverse wavelet transform subunit is used to perform an inverse wavelet transform on the processed first component, the processed second component, the processed third component, and the processed fourth component to obtain the subsequence for the k-th iteration and store the subsequence for the k-th iteration in the subsequence set.

[0171] The judgment subunit is used to judge whether the iteration end condition is satisfied to obtain a first judgment result. The iteration end condition is that k is equal to 3.

[0172] The output subunit is used to end the iteration and output the subsequences in the subsequence set if the first judgment result is yes.

[0173] The iteration subunit is used to subtract the subsequence for the k-th iteration from the building sequence for the k-th iteration to obtain the residual sequence for the k-th iteration if the first judgment result is no. Set the residual sequence for the k-th iteration as the building sequence for the k + 1-th iteration, increment the value of k by 1, and return to the decomposition subunit.

[0174] The permutation entropy value calculation module is used to calculate the permutation entropy value of each subsequence.

[0175] The prediction model determination module is used to determine the prediction model corresponding to each subsequence based on the permutation entropy value of each subsequence. The prediction model is an artificial neural network deep learning prediction model, a long short-term memory artificial neural network deep learning prediction model, a convolutional neural network-long short-term memory artificial neural network deep learning prediction model, or a temporal convolutional network model.

[0176] The prediction model determination unit is specifically used for:

[0177] Based on preset conditions and the permutation entropy values of each subsequence, determine the prediction model corresponding to each subsequence respectively; the preset conditions include using an artificial neural network model as the prediction model for the subsequence corresponding to the permutation entropy value that meets the first preset range, using a long short-term memory network model as the prediction model for the subsequence corresponding to the permutation entropy value that meets the second preset range, using a CNN-LSTM neural network model as the prediction model for the subsequence corresponding to the permutation entropy value that meets the third preset range, and using a temporal convolutional network model as the prediction model for the subsequence corresponding to the permutation entropy value that meets the fourth preset range.

[0178] A building energy consumption determination module, configured to obtain the building energy consumption based on each subsequence and the prediction model corresponding to each subsequence.

[0179] The building energy consumption determination module includes a prediction unit and a weighted summation unit. Now, the functions of each unit in the building energy consumption determination module are described:

[0180] The prediction unit is configured to input each subsequence into the prediction model corresponding to each subsequence to obtain the prediction result corresponding to each subsequence.

[0181] The weighted summation unit is configured to perform weighted summation on the prediction results corresponding to each subsequence to obtain the predicted building energy consumption.

[0182] An embodiment of the present application further provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the building energy consumption prediction method based on entropy evaluation.

[0183] An embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the building energy consumption prediction method based on entropy evaluation is implemented.

[0184] An embodiment of the present application further provides a computer program product, including a computer program, and when the computer program is executed by a processor, the building energy consumption prediction method based on entropy evaluation is implemented.

[0185] Although the existing hybrid prediction methods can effectively isolate the non-stationary information in the original data by combining traditional decomposition algorithms with prediction models, significantly improving the prediction accuracy. However, since traditional decomposition algorithms will generate many subsequences with different characteristics, it is difficult for a single prediction model to adapt to all subsequences, which brings a bottleneck to the overall prediction accuracy and efficiency. At the same time, the prediction of multiple decomposed subsequences will incur a large amount of time overhead, which will limit the application in scenarios with high timeliness requirements.

[0186] It should be noted that the various embodiments in this specification are described in a progressive manner. For the same or similar parts among the various embodiments, reference can be made to each other, and the key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the embodiments of the device, equipment, storage medium, and product, since they are basically similar to the method embodiments, the description is relatively simple, and reference can be made to the corresponding parts of the method embodiments for the relevant content. The device, equipment, storage medium, and product embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative efforts.

[0187] As described above, this is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A building energy consumption prediction method based on entropy assessment, characterized in that: The building energy consumption prediction method based on entropy assessment includes: Obtain building data; Decomposing the building data to obtain multiple subsequences; Calculate the permutation entropy value of each subsequence; Determine the prediction model corresponding to each subsequence based on the permutation entropy value of each subsequence; the prediction model is an artificial neural network deep learning prediction model, a long short-term memory artificial neural network deep learning prediction model, a convolutional neural network-long short-term memory artificial neural network deep learning prediction model or a time convolutional network model; The building energy consumption is obtained based on each subsequence and the prediction model corresponding to each subsequence.

2. The building energy consumption prediction method based on entropy assessment according to claim 1 is characterized in that: Decomposing the building data to obtain multiple subsequences specifically includes: Determine the wavelet basis function and decomposition threshold; The building data is decomposed three times based on the wavelet basis function and the decomposition threshold to obtain three subsequences.

3. The building energy consumption prediction method based on entropy assessment according to claim 2 is characterized in that: The three-time decomposition processing of the building data based on the wavelet basis function and the decomposition threshold to obtain three subsequences specifically includes: Let the value of k be 1; set the building data to the building sequence of the first iteration; Decomposing the k-th iteration building sequence by the wavelet basis function to obtain a component set; the component set includes a first component, a second component, a third component and a fourth component; the frequency of the first component is lower than the frequency of the second component; the frequency of the second component is lower than the frequency of the third component; the frequency of the third component is lower than the frequency of the fourth component; Setting the coefficients in the fourth component to 0, and performing soft thresholding processing on the coefficients in the first component, the second component, and the third component based on the decomposition threshold, respectively, to obtain a processed first component, a processed second component, a processed third component, and a processed fourth component; Performing inverse wavelet transform on the processed first component, the processed second component, the processed third component and the processed fourth component to obtain a subsequence of the kth iteration, and storing the subsequence of the kth iteration into a subsequence set; Determine whether an iteration end condition is met to obtain a first determination result; the iteration end condition is that k is equal to 3; If the first judgment result is yes, then the iteration ends and the subsequence in the subsequence set is output; If the first judgment result is no, subtracting the building sequence of the k-th iteration from the subsequence of the k-th iteration to obtain a residual sequence of the k-th iteration; The residual sequence of the k-th iteration is set as the building sequence of the k+1-th iteration, the value of k is increased by 1, and the step of decomposing the building sequence of the k-th iteration by the wavelet basis function to obtain a component set is returned.

4. The building energy consumption prediction method based on entropy assessment according to claim 1 is characterized in that: The determining the prediction model corresponding to each subsequence based on the permutation entropy value of each subsequence specifically includes: The prediction model corresponding to each subsequence is determined based on preset conditions and the permutation entropy value of each subsequence; the preset conditions include using an artificial neural network model as a prediction model for a subsequence corresponding to a permutation entropy value that meets a first preset range, using a long short-term memory network model as a prediction model for a subsequence corresponding to a permutation entropy value that meets a second preset range, using a CNN-LSTM neural network model as a prediction model for a subsequence corresponding to a permutation entropy value that meets a third preset range, and using a temporal convolutional network model as a prediction model for a subsequence corresponding to a permutation entropy value that meets a fourth preset range.

5. The building energy consumption prediction method based on entropy assessment according to claim 1 is characterized in that: The building energy consumption is obtained based on each subsequence and the prediction model corresponding to each subsequence, specifically including: Input each subsequence into the corresponding prediction model to obtain the prediction result corresponding to each subsequence; The predicted energy consumption of the building is obtained by weighted summing of the prediction results corresponding to each subsequence.

6. The building energy consumption prediction method based on entropy assessment according to claim 1 is characterized in that: The obtaining of building data specifically includes: Obtain building data to be processed; Determine whether there is an abnormal value in the building data to be processed, and obtain a second determination result; If the second judgment result is yes, the seasonal mixed ESD method is used to replace the abnormal value with the missing value, and the mean of the data before and after the missing value is calculated, and the missing value is replaced by the mean to obtain the building data; If the second judgment result is no, the building data to be processed is used as the building data.

7. A building energy consumption prediction device based on entropy assessment, characterized in that: The building energy consumption prediction device based on entropy assessment includes: An acquisition module, used to acquire building data; A decomposition module, used for decomposing the building data to obtain multiple subsequences; A permutation entropy value calculation module is used to calculate the permutation entropy value of each subsequence; A prediction model determination module, used to determine the prediction model corresponding to each subsequence based on the permutation entropy value of each subsequence; the prediction model is an artificial neural network deep learning prediction model, a long short-term memory artificial neural network deep learning prediction model, a convolutional neural network-long short-term memory artificial neural network deep learning prediction model or a time convolutional network model; The building energy consumption determination module is used to obtain the building energy consumption based on each subsequence and the prediction model corresponding to each subsequence.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the building energy consumption prediction method based on entropy assessment as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the building energy consumption prediction method based on entropy assessment described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the building energy consumption prediction method based on entropy assessment described in any one of claims 1 to 6 is implemented.

Images