A heating load prediction method based on similar day clustering and convolution gate recurrent

Abstract

The application discloses a heating load prediction method based on similar day clustering and convolution gate cycle, relates to the technical field of heating load prediction of central heat exchange station, and comprises the following steps: obtaining side heat supply load historical data and meteorological data, preprocessing the side heat supply load historical data and the meteorological data, slicing the side heat supply load historical data and the meteorological data according to days, and dividing the side heat supply load historical data and the meteorological data into a training set and a verification set; according to the change of the heat supply load, performing rough classification of weather types on the training set, performing fine classification on the weather types, and obtaining characteristics of different weather types; combining grey correlation degree analysis and Pearson product-moment correlation coefficient to obtain the maximum value of the correlation degree between a target day and each weather type, and selecting corresponding data sets as inputs; training through a combined neural network model, and obtaining a heat supply load prediction result. Therefore, the heating load prediction method based on similar day clustering and convolution gate cycle can improve the prediction accuracy of the heat load of the central heat exchange station and has stronger flexibility and universality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of heating load prediction of heat exchange station of centralized heat supply system, and particularly relates to a heating load prediction method based on similar day clustering and convolution gate cycle. BACKGROUND

[0002] Centralized heat supply system is an important part of people's livelihood project in northern China. The pipe network length of large centralized heat supply system usually extends for several kilometers, and it takes several hours or even longer for heat to reach the heat users from the heat source, which exhibits a prominent temperature transmission delay characteristic. This characteristic causes serious source-load mismatch and energy waste, reducing the economic and environmental benefits of the centralized heat supply system. Accurate prediction of the heat load of the heat exchange station node can provide reference for the change of the heat load for the operation and control personnel of the centralized heat supply system, and then adjust and optimize the subsequent operation strategy of the centralized heat supply system, so as to improve the source-load matching degree of the centralized heat supply system and improve the economic and environmental benefits of the centralized heat supply system.

[0003] Traditional heat load prediction research of centralized heat exchange system heat exchange station mostly adopts methods based on statistical theory. This method has strong interpretability and simple calculation, but has low prediction accuracy and poor robustness. With the rapid development and wide application of machine learning technology, it has become possible to accurately predict the heat load of each node in the heat exchange station using big data. For example, Xin Tan proposed a room heat load prediction calculation method based on the combination of analytic hierarchy process (AHP) and back propagation (BP) neural network, and took 8 related environmental factors as the input of the model, ensuring the accuracy and stability of the prediction results. Aliyu Ibrahim discussed the role of CNN in residential building load modeling, and through considering the shape characteristics of the building, one-dimensional data was predicted in multiple dimensions, which not only realized fast modeling, but also improved the prediction accuracy. Ling Z used a convolutional neural network and a random forest algorithm based on deep learning technology to establish a heat exchange station prediction model, and analyzed the data of a heat exchange station in Handan City throughout the heating season, improving the accuracy of heat load prediction of the heat exchange station. Hossein M F proposed and constructed a building heat load prediction model, and used real-time updated data to ensure the effectiveness of the information, which shows that artificial neural network can realize accurate prediction and has practical application value.

[0004] To further improve the prediction accuracy, some studies are devoted to exploring the method of multiple neural networks for heat load prediction. For example, a hybrid neural network is formed by combining CNN and LSTM, and further enhanced by combining wavelet decomposition for short-term heat load prediction. Compared with single neural network model, the hybrid neural network model has higher prediction accuracy, thereby better meeting the requirements of the heating system. On the other hand, some studies have improved the shortcomings of neural networks. For example, the combined neural network based on optimization method improves the accuracy of load energy consumption prediction calculation results; Junni S et al. introduced the adversarial neural network in the load prediction of integrated energy system, and used LSTM as the generator and discriminator, which effectively improved the prediction accuracy of the load; Yaohui H et al. used Ac-GRN (Active Graph Recurrent Network) to predict regional heat load, which effectively enhanced the interpretability of the prediction results.

[0005] Although the neural network method based on deep learning technology has great potential in the field of heat load prediction, it has a high requirement for the quality of the data set, and cannot capture long sequence features. Moreover, the heat load of the heat exchange station node is affected by many factors such as time, environment and human factors, and there is a mismatch between source and load in the central heating system and energy waste phenomenon, so it is necessary to provide a method based on environmental parameters and similar day clustering analysis to further improve the model prediction efficiency and accuracy and provide a reference for the operation of the central heating system. SUMMARY

[0006] The purpose of the present application is to provide a heating load prediction method based on similar day clustering and convolution gate cycle, which can improve the accuracy of heat load prediction of central heating system heat exchange station and has stronger flexibility and universality.

[0007] To achieve the above purpose, the present application provides a heating load prediction method based on similar day clustering and convolution gate cycle, comprising the following steps:

[0008] S1, obtain the side heat load historical data and meteorological data, and perform data preprocessing, then normalize the heat load and each meteorological parameter, slice by days, and divide the training set and the validation set;

[0009] S2, according to the change of heat load, first, the training set is roughly classified according to weather type, then the weather type is finely classified, and the characteristics of different weather types are obtained;

[0010] S3, according to the change of heat load under each weather type, combining gray correlation analysis and Pearson product-moment correlation coefficient, the correlation degree between the target day and each weather type is calculated, and the weather type data set with high correlation degree is selected as the input data set;

[0011] S4, based on the selected input data set, training and prediction are performed through a combined neural network model to obtain a prediction result of the target day heat load change.

[0012] Preferably, the coarse classification uses the DBSCAN clustering algorithm, and the fine classification uses the K-Means algorithm.

[0013] Preferably, the data preprocessing includes marking and filling the missing values in the data by using linear interpolation; detecting outliers by using the interquartile range of the box plot, setting them as null values, and then filling them by using linear interpolation; and smoothing the data by using an SG filter.

[0014] Preferably, the weather types are classified according to the change of the heat load, as follows:

[0015] H = f(x1, x2, ···, x F );

[0016] In the formula, H is the heat load, f represents the functional relationship between the environmental parameters and the heat load, x1, x2, ···, x F are the environmental parameters.

[0017] Preferably, the correlation between the target day and each weather type is calculated, including:

[0018] First, the Pearson product-moment correlation coefficients of the environmental parameters and the heat load under different weather types are calculated, and then the correlation between the target day and each weather type is obtained by gray correlation analysis, as follows:

[0019]

[0020] In the formula, ζ i (k) represents the gray correlation coefficient of the i-th meteorological factor of the k-th weather type, ρ represents the resolution coefficient, G min represents the global minimum value of the difference between the target day and the historical day, G max represents the global maximum value of the difference between the target day and the historical day, x i (k) represents the normalized value of the i-th meteorological factor of the k-th weather type, and x0(k) represents the normalized value of the corresponding meteorological factor of the target day.

[0021] Then, the Pearson product-moment correlation coefficients are used to weight and average the gray correlation coefficients of each meteorological factor under different weather types, thereby obtaining the correlation coefficients of the target day and each weather type.

[0022] Preferably, the combined neural network model includes a convolutional gated recurrent unit and a gated recurrent unit, and an attention mechanism is introduced.

[0023] Preferably, the convolutional gated recurrent unit extracts features by combining a convolutional neural network and a gated recurrent unit as follows:

[0024]

[0025] In the formula, σ represents a sigmoid activation function, W z , W r , and W represent weight matrices of update gates, reset gates, and candidate hidden layer states, z t , r t , represent outputs of update gates, reset gates, and candidate hidden layer states, h t , h t-1 respectively represent output values of state units at times t and t-1, and × represents a convolution operation.

[0026] Therefore, the heating load prediction method based on similar day clustering and convolutional gated recurrent units has the following technical effects:

[0027] (1) The training set is subjected to cluster analysis to obtain characteristics of different weather types, and then similar day analysis is used to obtain a data set with high correlation with the target day, thereby providing a high-precision data set for subsequent model training, preventing overfitting of the model, and enhancing robustness.

[0028] (2) The combined neural network can improve the feature extraction capability of the neural network, and the introduction of the attention mechanism improves the learning efficiency of the model, effectively improving the accuracy of heating load prediction of the central heating system heat exchange station.

[0029] The technical solutions of the present application will be described in further detail below with reference to the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a flowchart of a heating load prediction method based on similar day clustering and convolutional gated recurrent units;

[0031] Figure 2 is a Pearson product-moment correlation coefficient between each environmental parameter and the heating load in different weather types in an embodiment of the heating load prediction method based on similar day clustering and convolutional gated recurrent units;

[0032] Figure 3 is a grey correlation degree coefficient between each environmental parameter of the target day and different weather types in an embodiment of the heating load prediction method based on similar day clustering and convolutional gated recurrent units;

[0033] Figure 4A correlation coefficient between a target day and each weather type in an embodiment of a heating load prediction method based on similar day clustering and convolution gate cycle;

[0034] Figure 5 A combined neural network structure schematic diagram in an embodiment of a heating load prediction method based on similar day clustering and convolution gate cycle. DETAILED DESCRIPTION

[0035] The application can be explained in more detail by the following examples, and the purpose of disclosing the application is to protect all changes and improvements within the scope of the application, and the application is not limited to the following examples.

[0036] As Figure 1 shown, the application provides a heating load prediction method based on similar day clustering and convolution gate cycle, comprising the following steps:

[0037] Example 1

[0038] S1, in this embodiment, the heating load history data and meteorological data of the primary side of a heat exchange station from November 2022 to March 2023 are obtained, and the time step is 10 minutes.

[0039] Since there are missing values or abnormal values in the obtained original data, it is necessary to first preprocess them. In this embodiment, for the missing values in the data, linear interpolation is used to mark and fill the missing values; for the abnormal values, the box plot method is used to identify, that is, the quartile range of the box plot is used to detect abnormal points. After finding the abnormal values, set them to null values, and then linearly fill them.

[0040] After completing the processing of abnormal values and missing values, the SG filter (Savitzky Golay filter) is also used to smooth the data in this embodiment.

[0041] At the same time, since there are differences in dimensions between the heating load and each influencing parameter, it is necessary to normalize the heating load and each meteorological parameter, and slice them by day, and take the first 90% of the dates as the training set and the last 10% of the dates as the validation set.

[0042] S2, the size of the heating load is usually nonlinearly related to weather conditions such as ambient temperature, ambient humidity and other parameters, such as H = f(x1, x2, ···, x F ), where H is the heating load, f represents the functional relationship between each environmental parameter and the heating load, x1, x2, ···, x F is each environmental parameter. Therefore, the weather types can be classified according to the change of the heating load.

[0043] In this embodiment, first, the DBSCAN clustering algorithm is used to perform coarse classification of weather types according to the change of the heating load of the training set, and then the K-Means algorithm is used to perform fine classification, so as to obtain the change of the heating load under 10 weather types.

[0044] S3, to further show the difference of the heating load change under each weather type, in this embodiment, the Pearson product-moment correlation coefficient of meteorological factors and heating load is calculated, and it is found that the Pearson product-moment correlation coefficient between meteorological parameters and heating load is different for different weather types, as shown in the following formula: Figure 2

[0045] Taking a day in the verification set as the target day, the grey correlation degree analysis is used to obtain the correlation degree between the target day and each weather type, and the calculation formula is as follows:

[0046]

[0047] In the formula, ζ i (k) represents the grey correlation degree coefficient of the i th meteorological factor and the k th weather type, ρ represents the resolution coefficient, G min represents the global minimum value of the difference between the target day and the historical day, G max represents the global maximum value of the difference between the target day and the historical day, x i (k) represents the normalized value of the i th meteorological factor and the k th weather type, and x0(k) represents the normalized value of the corresponding meteorological factor of the target day.

[0048] Then, the Pearson product-moment correlation coefficient calculated is used to perform weighted average on the grey correlation degree coefficient of each meteorological factor under different weather types (ζ Figure 3 ), so as to obtain the correlation coefficient of the target day and each weather type, and the weather type data set with high correlation degree is selected as the input data set, as shown in the following formula: Figure 4

[0049] S4, based on the selected input data set, the combined neural network model is trained and predicted to obtain the prediction result of the heating load change of the target day.

[0050] As shown in the following formula, the combined neural network model includes a convolution gate recurrent unit and a gate recurrent unit, and an attention mechanism is introduced. The convolution gate recurrent unit performs feature extraction by combining a convolutional neural network and a gate recurrent unit, and the expression is as follows: Figure 5

[0051]

[0052] In the formula, σ represents a sigmoid activation function, W z , W​​​r , W represents the update gate, the reset gate, the weight matrix of the candidate hidden layer state, z t , r t , , h represents the output of the update gate, the reset gate, the candidate hidden layer state, h t , h t-1 , respectively, represent the output values of the state unit at times t and t-1, and x represents the convolution operation.

[0053] To verify the effectiveness of the similar day analysis method, in this embodiment, the data sets of each weather type are used as input to train and predict, and the accuracy is evaluated by mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), as shown in Table 1.

[0054] Table 1 Training accuracy of different weather type data sets

[0055] Weather type MSE (kW) RMSE (kW) MAE (kW) MAPE (%) Type 1 117.61 10.85 8.45 1.14 Type 2 238.20 15.43 12.22 1.65 Type 3 175.76 13.26 11.09 1.48 Type 4 124.49 11.16 8.28 1.11 Type 5 160.74 12.68 10.26 1.36 Type 6 220.99 14.87 12.72 1.72 Type 7 227.34 15.08 13.56 1.82 Type 8 360.10 18.98 16.32 2.21 Type 9 212.16 14.57 12.03 1.61 Type 10 160.75 12.68 10.57 1.41

[0056] As can be seen from Table 1, the data set corresponding to type 4 is used as input to train the combined neural network model, which performs best on each evaluation index, which is consistent with the results of Figure 3 The correlation coefficient between the target day and type 4 is the largest. This means that the method of the present embodiment can effectively select the input training set and improve the prediction accuracy and stability.

[0057] On the other hand, the present embodiment further trains (predicts heating load 1) using all data sets as input, trains (predicts heating load 2) using the existing similar day method, and trains (predicts heating load 3) using the method of the present embodiment, and analyzes the prediction results to verify the accuracy and reliability of the method of the present embodiment.

[0058] The prediction results obtained by training with different methods are shown in Table 2. It can be seen that, compared with the results of the two existing training methods, prediction of heating load 1 and prediction of heating load 2, the results of the method of the present embodiment reduce the MSE by 339.15 kW (27.61%), 140.16 kW (48.00%), respectively, the RMSE by 3.36 kW (14.91%), 7.41 kW (27.88%), respectively, the MAE by 2.58 kW (15.82%), 5.47 kW (28.49%), respectively, the MAPE by 0.34% (15.25%), 0.75% (28.41%), respectively, and the R 2 , respectively, by 0.12 (17.14%), 0.28 (40.00%).

[0059] Table 2 Accuracy of results obtained by training with different methods

[0060] Evaluation index Predicted heating load 1 Predicted heating load 2 Predicted heating load 3 MSE (kW) 706.63 507.64 367.48 RMSE (kW) 26.58 22.53 19.17 MAE (kW) 19.20 16.31 13.73 MAPE (%) 2.64 2.23 1.89 [R 2 ]]> 0.42 0.58 0.70

[0061] Example two

[0062] In order to verify the performance of the training data set selected by the method of the embodiment on different models, the existing CNN, LSTM, GRU and CMT models are used for comparative analysis, and the prediction results of different models are shown in Table 3.

[0063] As can be seen from Table 3, the fitting accuracy (R 2 ) of the improved neural network such as ConvGRU-GRU, CMT, PVT and CPVT is relatively high, and the prediction accuracy increases with the continuous optimization of the neural network. This shows that the data set corresponding to the weather type with high correlation degree with the target day selected by the method of the embodiment as the input of the neural network has higher flexibility and universality, effectively improves the prediction accuracy of the heat load of the heat exchange station in the central heating system, saves the calculation cost, and improves the practicability in engineering. Therefore, the neural network model can be optimized or replaced according to the actual demand.

[0064] Table 3: Result accuracy of different model training

[0065]

[0066] Therefore, the heating load prediction method based on similar day clustering and convolution gate cycle is used in the present application, which effectively improves the prediction accuracy of the heat load of the heat exchange station in the central heating system.

[0067] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application but not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: the technical solutions of the present application can still be modified or replaced by the equivalent, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A heating load prediction method based on similar day clustering and convolution gated recurrent, characterized in that, The method comprises the following steps: S1, obtaining side heat load historical data and meteorological data, and performing data preprocessing, then normalizing the heat load and each meteorological parameter, slicing by days, and dividing into a training set and a validation set; S2, according to the heat load change, first performing coarse classification of weather types on the training set, then performing fine classification of weather types, and obtaining characteristics of different weather types; S3, according to the heat load change under each weather type, combining gray correlation degree analysis and Pearson product-moment correlation coefficient, calculating the correlation degree between the target day and each weather type, and selecting the weather type data set with high correlation degree as the input data set, comprising: First, calculating the Pearson product-moment correlation coefficient between each environmental parameter and the heat load under different weather types, and then obtaining the correlation degree between the target day and each weather type through gray correlation degree analysis; wherein the Pearson product-moment correlation coefficient between the meteorological parameters and the heat load under different weather types is different; Then, using the calculated Pearson product-moment correlation coefficient, the gray correlation degree coefficients of each meteorological factor under different weather types are weighted and averaged, so as to obtain the correlation coefficient of the target day and each weather type; S4, based on the selected input data set, training and prediction are performed through a combined neural network model to obtain the prediction result of the heat load change of the target day; The combined neural network model comprises a convolutional gated recurrent unit and a gated recurrent unit, and an attention mechanism is introduced.

2. The heating load forecasting method based on similar day clustering and convolution gated recurrent according to claim 1, characterized in that, The data preprocessing comprises marking and filling the missing values in the data by using linear interpolation; detecting outliers by using the interquartile range of the box plot, setting the outliers as null values, and then filling the null values by using linear interpolation; and smoothing the data by using an SG filter.

3. The heating load forecasting method based on similar day clustering and convolution gated recurrent according to claim 1, characterized in that, The coarse classification adopts a DBSCAN clustering algorithm, and the fine classification adopts a K-Means algorithm.

4. The heating load forecasting method based on similar day clustering and convolution gated recurrent according to claim 1, characterized in that, The weather types are classified according to the heat load change, as follows: H = f(x1, x2, ···, x F ); In the formula, H is the heating load, f represents the functional relationship between each environmental parameter and the heating load, x1, x2, ···, x F are each environmental parameters.

5. The heating load forecasting method based on similar day clustering and convolution gated recurrent according to claim 1, characterized in that, The correlation degree between the target day and each weather type is calculated, comprising: First, calculating the Pearson product-moment correlation coefficient between each environmental parameter and the heat load under different weather types, and then obtaining the correlation degree between the target day and each weather type, as follows: wherein ζ i (k) represents the grey correlation degree coefficient of the i th meteorological factor for the k th weather type, p represents the resolution coefficient, G min represents the global minimum value of the difference between the target day and the historical day, G max represents the global maximum value of the difference between the target day and the historical day, x i (k) represents the normalized value of the i th meteorological factor for the k th weather type, x0(k) represents the normalized value of the target day corresponding to the meteorological factor; Then, using the calculated Pearson product-moment correlation coefficient, the gray correlation degree coefficients of each meteorological factor under different weather types are weighted and averaged, so as to obtain the correlation coefficient of the target day and each weather type.

6. The heating load forecasting method based on similar day clustering and convolution gated recurrent according to claim 1, characterized in that, The convolutional gated recurrent unit extracts features by combining a convolutional neural network and a gated recurrent unit, as follows: where σ denotes a sigmoid activation function, W z , W r , W denotes a weight matrix of the update gate, the reset gate, and the candidate hidden state, z t , r t , denotes the output of the update gate, the reset gate, and the candidate hidden state, h t , h t-1 denote the output values of the state cell at time t, t-1, respectively, and × denotes a convolution operation.

Images