Secondary network thermal load prediction method based on CEEMDAN-KPCA-LSTM algorithm

Abstract

The invention provides a secondary network thermal load prediction method based on a CEEMDAN-KPCA-LSTM algorithm. The method comprises the following steps: dividing a complemented data set into a training set and a test set according to a proportion; key factors influencing the dynamic change of the thermal load of the secondary network in the training set data are used as input variables, and the thermal load data are used as output variables; performing local encryption and standardization processing on the thermal load data; building an improved neural network model for predicting the dynamic thermal load of the building; training the improved neural network model, and when the improved neural network model meets a convergence condition, obtaining a trained improved neural network model; performing standardization processing on the test set data, and setting test set input data according to an input variable selected from the test set data after standardization processing of the training set data; and inputting test set input data into the trained improved neural network model to realize prediction of the dynamic thermal load of a certain building in the future 24 hours.

Description

Technical Field

[0001] The present invention belongs to the technical field of heat load prediction of heating systems, and in particular relates to a short-term heating heat load prediction method driven by historical data. Background Art

[0002] In towns in northern my country, winter heating mainly relies on centralized heating systems. In recent years, with the acceleration of urbanization, the heating area and heating capacity have increased year by year, which has led to a significant increase in energy consumption and carbon emissions. In the context of responding to the "dual carbon" goals, improving the accuracy of heat load forecasting and realizing on-demand heating have become one of the important ways to reduce heating energy consumption. According to different forecast periods, heat load forecasts can be divided into four types: ultra-short-term, short-term, medium-term and long-term. Among them, short-term heat load forecasts have the characteristics of strong periodicity, great influence of meteorological factors, and obvious load changes. This kind of forecast can help the heating system better adjust its operation strategy to adapt to real-time load changes, thereby improving heating efficiency and reducing energy waste.

[0003] Predictive control is a key strategy to improve the efficiency of unit building heating systems. It provides the best operation plan for the heating system by predicting the real-time heat load of the unit building. Under the current technological development background, heat load prediction methods are mainly divided into three types: white box method, black box method and gray box method. The white box method focuses on studying the physical mechanism of the unit building and predicts the heat load by analyzing the basic information of the unit building and heat transfer theory. However, since it requires a lot of unit building information, its universality is insufficient. The black box method is to make predictions by establishing a data-driven model and training and optimizing it based on a large amount of data. Although the relationship between the data is taken into account, it lacks an explanation of the physical mechanism of the unit building. The gray box method is a method that combines physical laws and data-driven models. On the basis of considering physical laws, it also makes full use of the advantages of data-driven models. Summary of the invention

[0004] In order to solve the problems existing in the prior art: the technical solution adopted in this application is as follows: a secondary network heat load prediction method based on CEEMDAN-KPCA-LSTM algorithm, characterized in that it includes the following steps:

[0005] Obtain the historical heating network data and historical environmental factor data of a building during the heating season to form a data set;

[0006] Clean the data set, remove outliers, and use the KNN algorithm to complete and preprocess the data set with missing data;

[0007] Divide the completed data set into training set and test set according to the proportion;

[0008] The key factors that affect the dynamic changes of the secondary network heat load in the training set data are used as input variables, and the heat load data is used as the output variable;

[0009] Locally encrypt and standardize heat load data;

[0010] Build an improved neural network model for predicting dynamic heat loads in buildings;

[0011] The improved neural network model is trained, and when the improved neural network model meets the convergence condition, a trained improved neural network model is obtained;

[0012] The test set data is standardized, and the test set input data is set according to the input variables selected from the standardized test set data by the training set data;

[0013] The test set input data is input into the trained improved neural network model to predict the future 24h dynamic heat load of a building.

[0014] Further: the improved neural network model includes fully integrating the empirical mode decomposition adaptive noise algorithm, kernel principal component analysis KPCA and long short-term memory neural network LSTM;

[0015] The fully integrated empirical mode decomposition adaptive noise CEEMDAN decomposes the environmental sequence into intrinsic mode components and residual components at different time scales, thereby increasing the diversity of features, and obtains the local features of environmental factors affecting heat load at different time scales through the complete set empirical mode decomposition of adaptive noise;

[0016] Then, kernel principal component analysis (KPCA) is used to map the high-dimensional data into a low-dimensional space by performing nonlinear mapping on the data, thereby extracting the main features of the data;

[0017] The LSTM neural network is used to perform dynamic time modeling of multivariate time series and heat load to achieve accurate prediction of the heating load of the heating system.

[0018] Further: The process of training the improved neural network model is as follows:

[0019] Determine the input data: environmental data a: outdoor temperature at time t-1, environmental data b: outdoor temperature at time t, environmental factor c: solar radiation at time t, environmental factor d: wind speed at time t, environmental factor e: heat load value at time t-1, environmental factor f: heat load value at time t-2;

[0020] The collected heat load data e, f and environmental data a, b, c, d are cleaned, and bad data in actual production are eliminated on a daily basis;

[0021] By fully integrating the empirical mode decomposition adaptive noise algorithm, the environmental data is decomposed into the intrinsic mode components {IMF1, IMF2, …, IMFm} of different frequencies and the residual component rn, and the original environmental sequence is decomposed into various characteristic fluctuation sequences to gradually decompose the fluctuations or trends of different scales existing in the original environmental signal;

[0022] The KPCA algorithm is used to screen out the key factors affecting the heat load from the decomposed environmental data, eliminating the redundancy and correlation of different time series data decomposed by CEEMDAN;

[0023] Normalize the reduced-dimensional data and heat load historical data, convert them into data sets suitable for LSTM network training, and divide them into training sets and test sets;

[0024] Initialize the parameters of the LSTM model and input the training set samples into the LSTM model for training until the measurement accuracy of the LSTM model is reached.

[0025] Further: the process of obtaining the local characteristics of the environmental factors affecting the heat load at different time scales through the complete set empirical mode decomposition of adaptive noise is as follows:

[0026] S11: The initial heat load data x of the heat exchange station after adding the adaptive white noise for the i-th time i (t) is written in the following form:

[0027] x i (t) = x(t) + ω 0 ε i (t),i∈{1,...,I} (2)

[0028] Where: x i (t) is the time series of heat load data with white noise added for the ith time; ω 0 is the noise coefficient; ε i (t) is the Gaussian white noise added for the i-th time; I is the number of integrations;

[0029] S12: Use EMD to calculate x i (t) is decomposed to obtain the intrinsic mode function IMF 1 and residual R 1 (t):

[0030]

[0031] R 1 (t) = x(t) - IMF 1 (t) (4)

[0032] S13: Use the signal x after adding noisei (t) minus IMF 1 (t) The residual signal is obtained, and the adaptive noise component is added to it. Then, EMD decomposition is performed to obtain the second component IMF 2 and residual R 2 (t):

[0033] S14: Execute step S13 again to obtain the k+1th modal component IMF K+1 (t) and the kth residual R K (t);

[0034] S15: When the decomposition termination condition is met, the decomposition ends, and the final result of the heat load data decomposition is:

[0035]

[0036] Where: x represents the input data, signal component IMF 2 and residual R 2 (t).

[0037] Further: the process of using the KPCA algorithm to screen out the key factors affecting the heat load from the decomposed environmental data is as follows:

[0038] Calculate the kernel matrix: First, you need to calculate the kernel matrix K of the original data set, where the output value of the kernel function is:

[0039] K ij =K(x i ,x j ) (17)

[0040] Centralized kernel matrix: Centralize the kernel matrix K by column or row to obtain the centralized kernel matrix Kc;

[0041] Solve eigenvalues ​​and eigenvectors: Perform eigendecomposition on the centralized kernel matrix K and solve its eigenvalues ​​and eigenvectors Kc;

[0042] Select principal components: Select the eigenvectors corresponding to the first k largest eigenvalues ​​as principal components to construct a new dimensionality reduction space;

[0043] Data projection: Project the original data set X into the new dimensionality reduction space to obtain the reduced dimensionality data matrix Y.

[0044] Further: the process of cleaning the data set, removing outliers, and using the KNN algorithm to complete the missing data set is as follows:

[0045] First, for missing data values, missing data for 12 consecutive hours are directly deleted. For other missing data, the K-NearestNeighbours algorithm is used to fill the missing values ​​by averaging the K neighboring samples around the missing samples to maintain the integrity of the data;

[0046] Secondly, for data outliers, abnormal data that appear repeatedly for 12 consecutive hours are directly deleted, and the Laida criterion is used to filter out outliers that are too large or too small, and the data group containing outliers is deleted.

[0047] Furthermore: the input variable selection selects key influencing factors of dynamic changes in heat load of the unit building through the heat transfer process of the unit building and the thermal characteristics of the unit building itself.

[0048] Furthermore: the method for selecting the key influencing factors includes but is not limited to Pearson correlation analysis, Spearman correlation analysis, Kendall rank, variance analysis, and principal factor analysis.

[0049] A secondary network heat load prediction device based on CEEMDAN-KPCA-LSTM algorithm, comprising:

[0050] Data acquisition module: used to obtain historical heating network data and historical environmental factor data of a building during the heating season;

[0051] Data preprocessing module: used to clean the data set, remove outliers, and use the KNN algorithm to complete the missing data set;

[0052] Data partitioning module: used to divide the completed data set into training set and test set according to the proportion;

[0053] Input variable selection module: used to take the key factors that affect the dynamic change of secondary network heat load in the training set data as input variables, and the heat load data as output variables;

[0054] Data encryption processing module: used for local encryption and standardization of heat load data;

[0055] Building module: used to build an improved neural network model for predicting dynamic heat loads of buildings;

[0056] Training module: used to train the improved neural network model. When the improved neural network model meets the convergence conditions, a trained improved neural network model is obtained.

[0057] Standardization processing module: used to perform standardization processing on the test set data, and set the test set input data according to the input variables selected from the standardized test set data by the training set data;

[0058] Implementation module: used to input the test set input data into the trained improved neural network model to predict the dynamic heat load of a building.

[0059] The invention discloses a secondary network heat load prediction method and device based on CEEMDAN-KPCA-LSTM algorithm, aiming to achieve system control strategy optimization and energy-saving operation. The model combines the fully integrated empirical mode decomposition adaptive noise method (CEEMDAN), kernel principal component analysis (KPCA) and long short-term memory neural network (LSTM). First, the environmental factor sequence is decomposed using CEEMDAN to reduce its non-stationarity; then, KPCA is used to extract the main influencing variables of the feature sequence to reduce the input dimension. In order to accurately predict the heating load of the heating system, the multivariate feature sequence is dynamically modeled using the LSTM network. This method is in response to the challenges of the operation and control of the current centralized heating system (especially the smart heating system). The secondary network heat load prediction method based on the CEEMDAN-KPCA-LSTM algorithm combines the thermal characteristics of the unit building, heat transfer theory and neural network technology, and can provide more accurate and efficient predictions. This method helps to improve the operation and control level of the heating system, thereby promoting the heating industry to achieve the goals of energy saving, emission reduction, efficiency improvement and quality improvement.

[0060] Compared with the prior art, the present invention has the following advantages:

[0061] 1. The secondary network heat load prediction method based on the CEEMDAN-KPCA-LSTM algorithm provided by the present invention determines the input variables of the heat load prediction model through physical mechanism and correlation analysis, so as to accurately extract the key factors affecting the dynamic thermal response of the building, thereby improving the prediction accuracy and the interpretability of the model.

[0062] 2. The secondary network heat load prediction method based on the CEEMDAN-KPCA-LSTM algorithm provided by the present invention is simple to implement and easy to understand. It can intuitively explain the prediction results of the building heat load, which is convenient for engineering application and practical operation.

[0063] 3. The secondary network heat load prediction method based on the CEEMDAN-KPCA-LSTM algorithm provided by the present invention is highly robust. It optimizes the loss function of the training process through physical mechanism, and proposes a physical loss function based on building room temperature, heat load and outdoor air temperature, which enhances the robustness of neural network training and improves the stability and reliability of the model.

[0064] 4. The secondary network heat load prediction method based on the CEEMDAN-KPCA-LSTM algorithm provided by the present invention better reveals the changing rules of data by locally encrypting the time step of data, and improves the model's adaptability to dynamic changes.

[0065] 5. The secondary network heat load prediction method based on the CEEMDAN-KPCA-LSTM algorithm provided by the present invention adopts the CEEMDAN algorithm for multi-scale decomposition and the KPCA algorithm for feature extraction, which effectively handles the nonlinear and non-stationary characteristics in the signal and improves the adaptability and prediction accuracy of the model.

[0066] 6. The high-precision prediction of the secondary network heat load prediction method based on the CEEMDAN-KPCA-LSTM algorithm provided by the present invention, combined with the three algorithms of fully integrated empirical mode decomposition adaptive noise method, kernel principal component analysis and long short-term memory network, can more accurately capture the complex change law of building heat load and improve the prediction precision and accuracy.

[0067] In view of the fact that the heat load of the heating system has the characteristics of instability and obvious seasonality, periodicity, intermittency, etc., this application proposes a heat load prediction model based on CEEMDAN-KPCA-LSTM, which fully considers the influence of environmental factors such as indoor and outdoor temperature, humidity, wind speed and solar radiation on the heat load. The model decomposes the environmental sequence into intrinsic mode components and residual components at different time scales through improved empirical mode decomposition (CEEMDAN), thereby increasing the diversity of features. It uses integrated learning and adaptive noise control technology on the basis of EMD, which can better handle the noise and pattern in the signal. Then, the kernel principal component analysis method (KPCA) is used to map the high-dimensional data to a low-dimensional space by nonlinear mapping of the data, thereby extracting the main features of the data. Unlike the traditional PCA method, KPCA can handle nonlinear relationships and can more accurately reflect the inherent structure of the data, eliminating the redundancy and correlation of different fluctuation sequences decomposed by CEEMDAN. Finally, the multivariate time series and heat load are dynamically modeled through the LSTM neural network to build a prediction model. Compared with the single LSTM model, EMD-LSTM model and EMD-KPCA-LSTM model, experimental results show that the model proposed in this paper has higher prediction accuracy.

[0068] In summary, through the research in this paper, the secondary network control strategy of the centralized heating system is optimized, which can realize on-demand heating and economical and energy-saving operation, and improve the heating effect. At the same time, the heat load prediction model based on CEEMDAN-KPCA-LSTM can predict the heat load more accurately, providing a reference for the operation and management of the heating system.

[0069] By combining CEEMDAN, KPCA and LSTM, the present invention can fully mine the information in the secondary network heat load data, and use deep learning technology for effective modeling and prediction. It has high prediction accuracy and practicality, and can be widely used in heating systems, thereby helping heating companies to improve heating quality, increase economic benefits, save energy, and reduce emissions of greenhouse gases and other pollutants.

[0070] Comprehensive advantages: By combining multiple methods and making full use of their respective advantages, a comprehensive prediction model is constructed, which can better cope with the challenges in dynamic heat load prediction of buildings and improve the overall performance and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0072] Figure 1 It is a flow chart of the present invention;

[0073] Figure 2 A flowchart of the improved neural network model of the present invention;

[0074] Figure 3 is the hidden layer cell structure of the LSTM of the present invention;

[0075] Figure 4 A comparison chart of heat load prediction results when the method is implemented using the improved LSTM algorithm provided by the present invention;

[0076] Figure 5 This is an error comparison chart of the heat load prediction results when the method is implemented using the improved LSTM algorithm provided by the present invention. DETAILED DESCRIPTION

[0077] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0078] Reference Figure 1As shown, the present invention discloses a secondary network heat load prediction method based on CEEMDAN-KPCA-LSTM algorithm, comprising the following steps:

[0079] S1: Obtain the historical heating network data and historical environmental factor data of a building during the heating season to form a data set;

[0080] S2: Clean the data set, remove outliers, and use the KNN algorithm to complete the missing data set;

[0081] S3: Divide the completed data set into a training set and a test set in a ratio of 7:3;

[0082] S4: The key factors that affect the dynamic change of the secondary network heat load in the training set data are used as input variables, and the heat load data is used as the output variable;

[0083] S5: Locally encrypt and standardize the heat load data; the local encryption is based on the changes in the input variables and heat load data, as well as factors such as sunrise and sunset time. After local encryption, the time step of the data is limited to a minimum of 15 minutes and a maximum of 1 hour;

[0084] S6: Build an improved neural network model for predicting dynamic heat load of buildings;

[0085] S7: training the improved neural network model, and when the improved neural network model meets the convergence condition, obtaining a trained improved neural network model;

[0086] S8: Standardize the test set data, and set the test set input data according to the input variables selected from the standardized test set data by the training set data;

[0087] S9: Input the test set input data into the trained improved neural network model to realize the prediction of the dynamic heat load of a building in the next 24 hours. During the prediction process, the heat load prediction result at the current moment is used as the basis for the heat load prediction at the next moment, thereby realizing multi-step prediction of the heat load.

[0088] Steps S1 / S2 / S3 / S4 / S5 / S6 / S7 / S8 / S9 are executed in sequence;

[0089] Data acquisition steps: Obtain historical heating network data and historical environmental factor data of a building during the heating season through actual measurement and numerical simulation, and construct a data set. The historical heating network data includes the data collection date and time period, the water supply flow, water supply temperature and return water temperature of each network involved in heating, and the historical environmental factor data includes outdoor temperature, solar radiation, wind speed, and humidity. The data in the data set mainly include building room temperature, outdoor temperature, solar radiation intensity, wind speed, heat load, and building surface area.

[0090] The correlation analysis of the candidate characteristic numerical data and the active power numerical data includes: judging the correlation between the candidate characteristic numerical data and the active power numerical data based on the Pearson correlation coefficient method, and the calculation formula is:

[0091]

[0092] Among them: |r|≤1 is the correlation coefficient between each factor; x i and i is the value of the two factors of the ith data point; and is the mean of the two factors; n is the number of data points.

[0093] In which, each characteristic numerical data set is regarded as a vector X=(x 1 ,x 2 ,...), the heat load data set is considered as a vector Y=(y 1 ,y 2 ,...), r is the measure of the similarity between two vectors, x i ,y i Represents the corresponding data value in the vector,

[0094] and Represents the mean of the vector elements; the strength of the correlation is determined by the size of r;

[0095] When 0.8≤r≤1.0, the candidate characteristic numerical data has a strong correlation with the heat load data;

[0096] When 0.6≤r<0.8, the candidate characteristic numerical data has a strong correlation with the heat load data;

[0097] When 0.4≤r<0.6, the candidate characteristic numerical data has a medium-strength correlation with the heat load data;

[0098] When 0.2≤r<0.4, the candidate characteristic numerical data has a weak correlation with the heat load data;

[0099] When 0≤r<0.2, the candidate characteristic numerical data has an extremely weak correlation with the heat load data.

[0100] Table 1 Correlation coefficients

[0101]

[0102] As can be seen from Table 1, the heat load has the highest positive correlation with the heat load at time t-1, the solar radiation at time t and the wind speed at time t are negatively correlated with the heat load data, and have a certain impact on the heat load output [7]. The outdoor temperature at time t has the greatest relative impact, which is consistent with expert common sense. Humidity has little effect on the heat load output. Therefore, the outdoor temperature at time t-1 and the outdoor temperature at time t, wind speed, solar radiation, and the heat load values ​​from time t-1 to t-2 are selected as model inputs.

[0103] Specifically, the value range of the Pearson correlation coefficient P is usually between -1 and 1. When P is close to 1, it indicates that there is a positive correlation between the candidate feature and the active power. When P is close to -1, it indicates that there is a negative correlation. When P is close to 0, it indicates that there is no linear correlation between the two.

[0104] According to the building heat transfer mechanism and the principle of variable correlation analysis, the Pearson, Kendall, and Spearman correlation indicators are used to select points with a greater correlation with the current heat load value from the room temperature, outdoor air temperature, solar radiation intensity, wind speed, heat load and other data in the training set as input variables for the prediction model. At the same time, considering the relationship between the building envelope area and the building heat transfer, the surface area of ​​the building envelope structure is selected as the input of the building heat load, and finally the input variable of the heat load LSTM prediction model is determined. The output variable is the heat load value.

[0105] Figure 2 A flowchart of the improved neural network model of the present invention;

[0106] The improved neural network model (CEEMDAN-KPCA-LSTM algorithm) includes fully integrated empirical mode decomposition adaptive noise method CEEMDAN), kernel principal component analysis KPCA and long short-term memory neural network LSTM;

[0107] The fully integrated empirical mode decomposition adaptive noise method CEEMDAN decomposes the environmental sequence into intrinsic mode components and residual components at different time scales, thereby increasing the diversity of features, and obtains the local characteristics of environmental factors affecting heat load at different time scales through the complete set empirical mode decomposition of adaptive noise; the CEEMDAN method is applied to perform empirical mode decomposition, decomposing the original data into multiple intrinsic mode functions (IMFs), thereby capturing the nonlinear and non-stationary characteristics in the data.

[0108] CEEMDAN (Complete Ensemble Empirical Mode Decomposition Adaptive Noise) is a signal processing method for decomposing complex signals into multiple intrinsic mode functions (IMFs) and a residual term. It is an improvement on the traditional empirical mode decomposition (EMD), mainly by introducing adaptive noise to solve the problem of modal aliasing and improve the accuracy of decomposition. This process involves adding adaptive noise to each decomposition, performing multiple decompositions and denoising, so as to obtain more stable and accurate decomposition results.

[0109] Then, kernel principal component analysis (KPCA) is used to map the high-dimensional data into a low-dimensional space by nonlinearly mapping the data, thereby extracting the main features of the data; KPCA dimensionality reduction processing is performed on the IMF decomposed by CEEMDAN to map the high-dimensional data into a low-dimensional space to reduce the complexity and redundancy of the data and improve the efficiency of subsequent processing. KPCA is used to extract features from the IMF decomposed by CEEMDAN. KPCA is a nonlinear dimensionality reduction technique that discovers the nonlinear structure of the data by mapping the data into a high-dimensional space and calculating the principal components in the space.

[0110] The LSTM (LSTM is a special type of recurrent neural network (RNN) used to process and predict sequence data, and is particularly suitable for capturing long-term dependencies) neural network is used to perform dynamic time modeling of multivariate time series and heat load to achieve accurate prediction of the heating load of the heating system.

[0111] Training the improved neural network model includes:

[0112] The prediction accuracy of the improved neural network model is evaluated, and necessary parameter and structure tuning is performed.

[0113] Training end judgment step: judge whether the error analysis reaches the preset threshold value. If not, return to the neural network prediction model training step; if yes, end the training and output the neural network prediction model;

[0114] Implement LSTM algorithm training. Apply the output variables and selected input variables to the LSTM algorithm for algorithm training and perform parameter tuning through cross-validation and other techniques to improve the performance and generalization ability of the model. Perform data standardization before training. The training loss function is the conventional loss function L original +(Physical loss factor λ phy ×Physical loss function L phy). The conventional loss function is the root mean square error between the predicted value and the actual value of the heat load. The physical loss factor is taken as 1. The physical loss function is the root mean square error between the predicted value and the actual value of the room temperature.

[0115] Further: the data set in the step of acquiring data includes but is not limited to indoor temperature data of the unit building, outdoor environmental status data, heating system operation data and unit building body information data.

[0116] Further: the actual value in the training step of the improved neural network prediction model is composed of the historical heat load data of the unit building.

[0117] Further: Fully Integrated Empirical Mode Decomposition Adaptive Noise is a data processing method used to decompose the original non-stationary signal into multiple essential mode functions (IMFs), each of which represents the oscillation components of different scales in the signal. CEEMDAN can adaptively decompose the local characteristics of the signal and can effectively handle the nonlinearity and non-stationarity in the signal, which is particularly important for processing the complex fluctuation characteristics in the unit building energy consumption data.

[0118] Further: KPCA is a principal component analysis technique based on the kernel method, which can effectively handle nonlinear data relationships and extract the main features in the data. Applying KPCA on the IMFs obtained after CEEMDAN decomposition can further extract high-order statistical features in the data and map them to a higher-dimensional feature space, thereby better capturing the complex structure and dynamic changes of the data.

[0119] Further: LSTM is a deep learning model suitable for modeling sequence data. It has memory units and gating mechanisms, which can effectively capture long-term dependencies and time series patterns in data. Applying LSTM on the high-order features extracted by KPCA can build an end-to-end deep learning model that can learn the dynamic changes of heat load from historical data and use these laws for future predictions.

[0120] Furthermore: in the input variable selection step, key influencing factors of the dynamic change of the unit building heat load are selected through the unit building heat transfer process and the unit building body thermal characteristics.

[0121] Furthermore: methods for selecting key influencing factors include but are not limited to Pearson correlation analysis, Spearman correlation analysis, Kendall rank, variance analysis, and principal factor analysis.

[0122] Furthermore: the principle of local encryption of heat load data covers the changing trends of key influencing factors and the changes in the heat load data of the unit building itself.

[0123] Further: In the process of dynamic heat load prediction of unit buildings, one of the important links is model evaluation and tuning, as well as error verification. In this stage, we use the neural network prediction model to predict the test data set. Then, the prediction results are compared with the actual output variables in the test data set to measure the difference between the predicted values ​​and the actual observed values, and then evaluate the accuracy of the model. By adjusting the model parameters or using other optimization methods, we try to improve the prediction performance of the model to better meet the actual application needs.

[0124] The schematic diagram of the CEEMDAN-KPCA-LSTM algorithm implementing this method. The implementation steps are as follows:

[0125] 1) Collect data of a building through actual measurement and numerical simulation to construct a data set. The data in the data set mainly include building room temperature, outdoor temperature, solar radiation intensity, wind speed, heat load, and building surface area. Divide the data set into a training set and a test set. The training set is 70% of the data in the data set. The test set is 30% of the data in the data set. Remove outliers from the data set and use the KNN algorithm to obtain missing points.

[0126] 2) According to the principle of building heat transfer mechanism and variable correlation analysis, through Pearson, Kendall, and Spearman correlation indicators, in the training set, the points with greater correlation with the current heat load value are selected from the room temperature, outdoor air temperature, solar radiation intensity, wind speed, heat load and other data as the input variables of the prediction model. At the same time, considering the relationship between the building envelope area and the building heat transfer, the surface area of ​​the building envelope structure is selected as the input of the building heat load, and finally the input variable of the heat load LSTM prediction model is determined. The output variable is the heat load value.

[0127] 3) CEEMDAN decomposes the original signal into multiple IMFs and a residual term. These IMFs can capture the local characteristics of the signal and obtain a new data feature sequence.

[0128] 4) Kernel Principal Component Analysis (KPCA) is used to extract features from the IMF decomposed by CEEMDAN. KPCA is a nonlinear dimensionality reduction technique that discovers the nonlinear structure of data by mapping data into a high-dimensional space and calculating the principal components in the space.

[0129] 5) The data is locally encrypted based on the changes in the input variables and heat load data, as well as factors such as sunrise and sunset time. After local encryption, the time step of the data is limited to a minimum of 15 minutes and a maximum of 1 hour.

[0130] 6) Implement LSTM algorithm training. Apply the output variables and selected input variables to the LSTM algorithm for algorithm training and perform parameter tuning through cross-validation and other techniques to improve the performance and generalization ability of the model. Perform data standardization before training. The training loss function is the conventional loss function L original +(Physical loss factor λ phy ×Physical loss function L phy ). The conventional loss function is the root mean square error between the predicted value and the actual value of the heat load. The physical loss factor is taken as 1. The physical loss function is the root mean square error between the predicted value and the actual value of the room temperature.

[0131] 7) Use the trained CEEMDAN-KPCA-LSTM model to predict future data to predict the heat load value for the next 24 hours. During the prediction process, the heat load prediction result at the current moment is used as the basis for the heat load prediction at the next moment, thereby realizing multi-step prediction of the heat load.

[0132] This application uses the CEEMDAN method and the KPCA method to process various environmental factor data. The local characteristics of environmental factors affecting heat load at different time scales can be obtained through the complete set empirical mode decomposition of adaptive noise, which will obtain more detailed but more influencing factors. Then the kernel principal component analysis method is used for dimensionality reduction, and the feature dimension is reduced as the input of the long short-term memory neural network without losing information, so as to improve the computational efficiency and accuracy of the deep neural network.

[0133] The process of complete ensemble empirical mode decomposition with adaptive noise is as follows:

[0134] Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is a decomposition algorithm proposed by Torres in 2011. The basic idea of ​​the CEEMDAN algorithm is to eliminate the reconstruction error by adaptively adding zero total Gaussian white noise to the original data at each decomposition stage. When the decomposition error is almost zero, this method can solve the modal mixing problem of EMD and solve the false component problem in EEMD by using finite integral calculation.

[0135] The following is the complete ensemble empirical mode decomposition process for adaptive noise:

[0136] S11: Initial heat load data x of the heat exchange station after adding adaptive white noise for the i-th time i (t) can be written in the following form:

[0137] x i(t) = x(t) + ω 0 ε i (t),i∈{1,...,I} (2)

[0138] Where: x i (t) is the time series of heat load data with white noise added for the ith time; ω 0 is the noise coefficient; ε i (t) is the Gaussian white noise added for the i-th time; I is the number of integrations;

[0139] S12: Using EMD to calculate x i (t) is decomposed to obtain IMF 1 and residual R 1 (t):

[0140]

[0141] R 1 (t) = x(t) - IMF 1 (t) (4)

[0142] S13: Use the signal x after adding noise i (t) minus IMF 1 (t) The remaining signal can be obtained by adding the adaptive noise component to it. Then, the EMD decomposition is performed to obtain the second component IMF 2 and residual R 2 (t):

[0143] S14: Execute step S13 again to obtain the k+1th modal component IMF K+1 (t) and the kth residual R K (t):

[0144] S15: When the decomposition termination condition is met, the decomposition ends. The final result of the heat load data decomposition is:

[0145]

[0146] The data sequence obtained by CEEMDAN decomposition enriches the number of feature sequences, but the dimension of the input variables also increases. In order to improve the prediction accuracy while maintaining the calculation speed of the LSTM network model and overcome the problem of overfitting, it is necessary to reduce the dimension of the input variables through PCA, and improve the calculation efficiency and accuracy of the model while ensuring the validity and representativeness of the information.

[0147] The decomposition process can be terminated when a preset number of modal functions is reached. Generally speaking, when the signal is decomposed into a sufficient number of IMFs, the remaining signal part is considered to be a trend term or a noise term, and the decomposition can be stopped at this time.

[0148] Nonlinear and separable data can be processed using kernel-based principal component analysis (KPCA). The KPCA method maps the original data into a high-dimensional feature space and then uses PCA to reduce the dimensionality of the data, thereby enhancing the linear separability of the data in this space.

[0149] The kernel function is used to calculate the similarity between two samples. Its output is the similarity value between the two samples received as input. The kernel function of the KPCA algorithm maps the data in the original space to a new high-dimensional space, thereby improving the linear separability of the data in the new space.

[0150] Gaussian radial basis function kernel function:

[0151]

[0152] Among them, x i and x j Represents the input sample, and σ is the parameter that controls the shape of the kernel function.

[0153] The LSTM network model is a type of deep neural network. On the basis of the recurrent neural network (RNN), the forget gate, input gate and output gate are added to solve the problems of RNN gradient explosion and gradient disappearance, and enable the model to selectively remember relevant information and eliminate irrelevant information. The LSTM network model can learn long-term dependent information and achieve the "memory" function by inferring the state of the next moment through the state of the previous moment, thereby effectively improving the prediction accuracy of the model.

[0154] For a given sequence x=(x 1 ,x 2 ,...,x n ), the iterative calculation of the RNN structure is shown as follows:

[0155] h t =f a (W xh x t +W hh h t-1 +b h ) (7)

[0156] y t =W hy h t +b y (8)

[0157] Where: W is the weight coefficient matrix; b is the bias vector; f a is the activation function; subscript t is the time.

[0158] The hidden layer cell structure of LSTM is as follows Figure 3 As shown in the figure, by introducing the control gate, the problem of RNN gradient disappearance and explosion is solved, making the deep neural network unfolded in time easier to train. The forward calculation process is shown in the following formula:

[0159] i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i ) (9)

[0160] f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f ) (10)

[0161] c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c ) (11)

[0162] o t =σ(W xo x t +W ho h t-1 +W co c t +b o ) (12)

[0163] In the formula: i——input gate; f——forget gate; c——cell state; o——output gate; W——corresponding weight coefficient matrix; b——corresponding bias term; σ——sigmoid activation function; tanh——hyperbolic tangent activation function.

[0164] This application constructs a heat load prediction model of CEEMDAN-KPCA-LSTM. In the actual prediction process, a data set suitable for LSTM model training is first constructed, and the heat load at time t is predicted using the environmental factor data at time t and the heat load data at time t-1.

[0165] The improved neural network model is trained. When the improved neural network model meets the convergence conditions, the process of obtaining the trained improved neural network model is as follows:

[0166] Input: Environmental data a: outdoor temperature at time t-1, environmental data b: outdoor temperature at time t, environmental data c: solar radiation at time t, environmental data d: wind speed at time t, environmental data e: heat load value at time t-1, data f - heat load value at time t-2.

[0167] Output: model evaluation indicators MAPE, RMSE, R 2 .

[0168] 1) Data cleaning: The collected heat load data e, f and environmental data a, b, c, d are cleaned, and the "bad data" caused by communication failures and other reasons in actual production are eliminated on a daily basis;

[0169] 2) Through the CEEMDAN algorithm, the environmental data are decomposed into intrinsic mode components {IMF1, IMF2, …, IMFm} of different frequencies and residual components rn, and the original environmental sequence is decomposed into various characteristic fluctuation sequences to gradually decompose the different scale fluctuations or trends existing in the original environmental signal.

[0170] 3) Perform KPCA dimensionality reduction on the data decomposed in step 2), use the KPCA algorithm to screen out the key factors affecting the heat load, and eliminate the redundancy and correlation of different time series data decomposed by CEEMDAN.

[0171] 4) The data after dimension reduction in step 3) and the historical data of heat load are normalized and converted into a data set suitable for LSTM network training, and divided into training set and test set.

[0172] 5) Initialize the parameters of the LSTM model and input the training set samples into the LSTM model for training until the measurement accuracy of the model is achieved.

[0173] 6) After the training of the improved neural network model is completed, save the training file and input the test set for testing.

[0174] 7) Output the improved neural network model evaluation indicators MAPE, RMSE, R 2 , the algorithm ends.

[0175] For the short-term prediction model of heat exchange station that has been established, it is necessary to clarify the accuracy of the current model in the prediction after the prediction. Therefore, it is necessary to establish a comprehensive evaluation index suitable for the model itself to test the effectiveness of the prediction model. Heat load prediction is one of the regression problems. The evaluation indexes commonly used for regression problems are as follows:

[0176]

[0177] Among them: root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), coefficient of determination (R^2), Y i is the ith observed value or actual value; y i is the i-th predicted value, and P is the total number of observed data;

[0178] This application uses RMSE, MAE, MAPE and R 2 Together they serve as indicators for evaluating the accuracy of heat load prediction schemes.

[0179] Figure 4 and Figure 5 They are comparison charts of the prediction results of four models and comparison charts of the errors of the prediction results of four models, which show that the improved neural network model can make more accurate predictions.

[0180] Example 1: The experimental sample data used are the actual heating data from November 27, 2023 to February 27, 2024 provided by a thermal power group in Dalian.

[0181] S1: Preprocessing of experimental data;

[0182] In actual projects, since the monitoring equipment is affected by realistic factors such as network signal fluctuations and power outages, the monitoring data may be missing or abnormal. In order to improve the overall quality of the input data, it is necessary to process the original data. Python is used to process 1659 sets of hourly data collected from a residential community during the heating season from 2023 to 2024.

[0183] First, we used two methods to deal with missing data values. Missing data for 12 consecutive hours were directly deleted because these data could not accurately reflect the operation of the heating system. For other missing data, the K-Nearest Neighbours (KNN) algorithm was used to fill in the missing values. This algorithm fills in the missing values ​​by averaging the K neighboring samples around the missing samples to maintain the integrity of the data.

[0184] Secondly, we also adopted two methods to deal with data outliers. Abnormal data that appeared repeatedly for 12 consecutive hours was directly deleted, because such data was likely to be mismeasured due to equipment failure or other problems. At the same time, we used the Laida criterion to filter out outliers that were too large or too small, and deleted the data groups containing outliers, which ensured the accuracy and reliability of the data. After the above processing of missing values ​​and outliers, this paper finally obtained 1,348 sets of valid data. After processing, these data can better reflect the operation of the heating system in residential areas and provide a reliable basis for subsequent analysis and modeling.

[0185] In order to further process the data, we also standardized the data. Standardization can make the data have the same scale, which is conducive to the agreement of the model and the improvement of performance. We standardized the original data to make it conform to the standard normal distribution or within a certain range. Then, we used 70% of the data as the training set and the rest of the data as the test set for subsequent modeling and analysis.

[0186] S2: CEEMDAN decomposition results;

[0187] The environmental sequence data in the experimental samples are non-stationary signals, which are greatly affected by weather changes and have certain randomness and mutation. This paper uses the CEEMDAN algorithm to decompose the original environmental sequence data to obtain the IMF components and residual components in each environmental factor data, thereby highlighting the local characteristics of the original environmental sequence. After CEEMDAN decomposition of each data sequence, 56-dimensional IMF components and 7-dimensional residual components are obtained, which contain a total of 64-dimensional feature sequences as a new feature sequence set.

[0188] S3: KPCA dimensionality reduction;

[0189] In order to remove the noise in the feature sequence data and reduce the redundancy and correlation of the feature sequence data, kernel principal component analysis is performed on the obtained feature sequence data.

[0190] The steps of KPCA algorithm are as follows:

[0191] Calculate the kernel matrix: First, you need to calculate the kernel matrix K of the original data set, where the output value of the kernel function is:

[0192] K ij =K(x i ,x j ) (17)

[0193] Centralized kernel matrix: The kernel matrix K is centralized by columns or rows to obtain the centralized kernel matrix Kc.

[0194] Solve for eigenvalues ​​and eigenvectors: Perform eigendecomposition on the centralized kernel matrix K and solve its eigenvalues ​​and eigenvectors Kc.

[0195] Select principal components: Select the eigenvectors corresponding to the first k largest eigenvalues ​​as principal components to construct a new dimensionality reduction space.

[0196] Data projection: Project the original data set X into the new dimensionality reduction space to obtain the reduced dimensionality data matrix Y.

[0197] In the 64-dimensional feature sequence after CEEMDAN dimension reduction, the eigenvalues ​​of the first 10 feature sequences are all greater than, and the cumulative contribution rate reaches 90%. The selected principal components have a strong representativeness of the original feature sequence and a stronger ability to integrate information. Therefore, the first 10 feature sequences are selected to replace the original variables. Finally, the values ​​of the first 10 principal components are calculated. Through KPCA dimensionality reduction, the dimension and complexity of data input are reduced, and the input sample of the heat load prediction model at a certain moment in the future is constructed.

[0198] S4 improved LSTM model parameters

[0199] The 1659 time section data in the sample data set are converted into a data set suitable for LSTM network training, and are divided into training set and test set in a ratio of 7:3. The adaptive momentum estimation algorithm is used to optimize the update weights in the training process. The model parameters required to establish the LSTM network for heat load prediction are debugged as follows: 1) the number of input layer time steps is 1; 2) the input layer dimension is 10; 3) the number of hidden layers is 1; 4) the number of hidden layer units is 50; 5) the output layer variable dimension is 1; 6) the training batch is 10; 7) the training round is 100; 8) the training set data is 1161 groups; 9) the test set data is 498 groups.

[0200] Experimental results analysis

[0201] In order to verify the superiority of the CEEMDAN-KPCA-LSTM heat load prediction model proposed in this paper, the effectiveness of CEEMDAN decomposition and the necessity of KPCA dimensionality reduction, and further illustrate that the proposed method has a significant improvement in the prediction of heating load, four prediction models, namely a single LSTM model, an EMD-LSTM model, an EMD-KPCA-LSTM model and a CEEMDAN-KPCA-LSTM model, are used to predict the heat load in the next 24 hours. In order to more clearly demonstrate the prediction effect of the model in this paper, the prediction data can be compared with the real data and the four model evaluation indicators can be compared to verify the accuracy of the CEEMDAN-KPCA-LSTM heat load prediction proposed in this paper, and the prediction effect can be shown in a graph.

[0202] Table 2 Model prediction evaluation

[0203]

[0204]

[0205] As shown in Table 2, the CEEMDAN-KPCA-LSTM model performs best in terms of RMSE and MAE, which are 7.53 and 6.76 respectively, which are smaller than other models, indicating that the model has higher prediction accuracy. In terms of MAPE (mean absolute percentage error), the CEEMDAN-KPCA-LSTM model also performs best, which is 1.63%, indicating that its average prediction error is relatively small. In addition, the CEEMDAN-KPCA-LSTM model performs best in R 2 The goodness of fit is also good, at 98.34%, which means that the model can fit the changing trend of the actual data well.

[0206] In view of the large volatility and randomness of heat load output values, this application proposes a hybrid prediction method for heat load prediction based on CEEMDAN-KPCA-LSTM neural network, which fully considers the main factors restricting heat load, namely, outdoor temperature at time t-1 and outdoor temperature at time t, wind speed, solar radiation, and heat load value from time t-1 to time t-2. The main conclusions are as follows:

[0207] 1) The environmental sequence is decomposed using the complete set empirical mode decomposition method of adaptive noise to obtain different intrinsic mode components and residual terms. The original environmental sequence can be decomposed into various characteristic fluctuation sequences, thereby decomposing the fluctuations or trends of different scales in the original environmental signal step by step.

[0208] 2) Using the kernel principal component analysis method to screen out the key influencing factors affecting the heat load is a necessary condition to reduce the dimensionality of the model input parameters and eliminate the redundancy and correlation of different time series obtained by CEEMDAN decomposition.

[0209] 3) The LSTM neural network is used to perform dynamic time modeling on the nonlinear relationship between the multivariate time series and the heat load series, and a prediction model is constructed to finally complete the prediction of the heat load. The example analysis shows that compared with the LSTM model, the CEEMDAN-KPCA-LSTM model has a RMSE reduction of about 15.97%, a MAE reduction of about 3.71%, a MAPE reduction of about 44.24%, and a R 2 Compared with the EMD-LSTM model, the CEEMDAN-KPCA-LSTM model has a decrease in RMSE by about 41.95%, MAE by about 38.36%, and MAPE by about 65.68%. 2Compared with the EMD-KPCA-LSTM model, the CEEMDAN-KPCA-LSTM model has a decrease in RMSE by about 10.5%, MAE by about 5.94%, and MAPE by about 55.86%. 2 It increased by about 0.93%. In general, the CEEMDAN-KPCA-LSTM model showed better performance than other models in various indicators.

[0210] This application verifies the practicality of the LSTM model in the field of heating and expands the application scope of deep learning technology. The proposed prediction method provides a new perspective for in-depth exploration of the economic operation and scheduling of heating systems. It has good application prospects and engineering application value in reality, and related research is still in progress.

[0211] The experimental results show that the prediction accuracy of the model is higher than that of the traditional method when verifying the measured data of a heat exchange station in Dalian. This method fully considers the impact of environmental factors on heat load and provides important theoretical support and practical guidance for achieving energy-saving operation of heating systems.

[0212] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without creative work.

[0213] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A secondary network heat load prediction method based on CEEMDAN-KPCA-LSTM algorithm, characterized by: The following steps are involved: Obtain the historical heating network data and historical environmental factor data of a building during the heating season to form a data set; Clean the data set, remove outliers, and use the KNN algorithm to complete and preprocess the data set with missing data; Divide the completed data set into training set and test set according to the proportion; The key factors that affect the dynamic changes of the secondary network heat load in the training set data are used as input variables, and the heat load data is used as the output variable; Locally encrypt and standardize heat load data; Build an improved neural network model for predicting dynamic heat loads in buildings; The improved neural network model is trained, and when the improved neural network model meets the convergence condition, a trained improved neural network model is obtained; The test set data is standardized, and the test set input data is set according to the input variables selected from the standardized test set data by the training set data; The test set input data is input into the trained improved neural network model to predict the future 24h dynamic heat load of a building.

2. The secondary network heat load prediction method based on CEEMDAN-KPCA-LSTM algorithm according to claim 1 is characterized in that: The improved neural network model includes fully integrating the empirical mode decomposition adaptive noise algorithm, kernel principal component analysis KPCA and long short-term memory neural network LSTM; The fully integrated empirical mode decomposition adaptive noise CEEMDAN decomposes the environmental sequence into intrinsic mode components and residual components at different time scales, thereby increasing the diversity of features, and obtains the local features of environmental factors affecting heat load at different time scales through the complete set empirical mode decomposition of adaptive noise; Then, kernel principal component analysis (KPCA) is used to map the high-dimensional data into a low-dimensional space by performing nonlinear mapping on the data, thereby extracting the main features of the data; The LSTM neural network is used to perform dynamic time modeling of multivariate time series and heat load to achieve accurate prediction of the heating load of the heating system.

3. The secondary network heat load prediction method based on CEEMDAN-KPCA-LSTM algorithm according to claim 1 is characterized in that: The process of training the improved neural network model is as follows: Determine the input data: environmental data a: outdoor temperature at time t-1, environmental data b: outdoor temperature at time t, environmental factor c: solar radiation at time t, environmental factor d: wind speed at time t, environmental factor e: heat load value at time t-1, environmental factor f: heat load value at time t-2; The collected heat load data e, f and environmental data a, b, c, d are cleaned, and bad data in actual production are eliminated on a daily basis; By fully integrating the empirical mode decomposition adaptive noise algorithm, the environmental data is decomposed into the intrinsic mode components {IMF1, IMF2, …, IMFm} of different frequencies and the residual component rn, and the original environmental sequence is decomposed into various characteristic fluctuation sequences to gradually decompose the fluctuations or trends of different scales existing in the original environmental signal; The KPCA algorithm is used to screen out the key factors affecting the heat load from the decomposed environmental data, eliminating the redundancy and correlation of different time series data decomposed by CEEMDAN; Normalize the reduced-dimensional data and heat load historical data, convert them into data sets suitable for LSTM network training, and divide them into training sets and test sets; Initialize the parameters of the LSTM model and input the training set samples into the LSTM model for training until the measurement accuracy of the LSTM model is reached.

4. The secondary network heat load prediction method based on CEEMDAN-KPCA-LSTM algorithm according to claim 1 is characterized in that: The process of obtaining the local characteristics of environmental factors affecting heat load at different time scales through the complete set empirical mode decomposition of adaptive noise is as follows: S11: The initial heat load data x of the heat exchange station after adding the adaptive white noise for the i-th time i (t) is written in the following form: x i (t)=x(t)+ω0ε i (t),i∈{1,...,I} (2) Where: x i (t) is the time series of heat load data with white noise added for the ith time; ω0 is the noise coefficient; ε i (t) is the Gaussian white noise added for the i-th time; I is the number of integrations; S12: Use EMD to calculate x i (t) is decomposed to obtain the intrinsic mode function IMF1 and residual R1(t): R1(t)=x(t)-IMF1(t) (4) S13: Use the signal x after adding noise i (t) minus IMF1(t) to get the residual signal, to which the adaptive noise component is added Then, EMD decomposition is performed to obtain the second component IMF2 and the residual R2(t): S14: Execute step S13 again to obtain the k+1th modal component IMF K+1 (t) and the kth residual R K (t); S15: When the decomposition termination condition is met, the decomposition ends, and the final result of the heat load data decomposition is: Where: x represents the input data, the signal component IMF2 and the residual R2(t).

5. The secondary network heat load prediction method based on CEEMDAN-KPCA-LSTM algorithm according to claim 1 is characterized in that: The process of using the KPCA algorithm to screen out the key factors affecting the heat load from the decomposed environmental data is as follows: Calculate the kernel matrix: First, you need to calculate the kernel matrix K of the original data set, where the output value of the kernel function is: K ij =K(x i ,x j ) (17) Centralized kernel matrix: Centralize the kernel matrix K by column or row to obtain the centralized kernel matrix Kc; Solve eigenvalues ​​and eigenvectors: Perform eigendecomposition on the centralized kernel matrix K and solve its eigenvalues ​​and eigenvectors Kc; Select principal components: Select the eigenvectors corresponding to the first k largest eigenvalues ​​as principal components to construct a new dimensionality reduction space; Data projection: Project the original data set X into the new dimensionality reduction space to obtain the reduced dimensionality data matrix Y.

6. The secondary network heat load prediction method based on CEEMDAN-KPCA-LSTM algorithm according to claim 1 is characterized in that: The process of cleaning the data set, removing outliers, and using the KNN algorithm to complete the missing data set is as follows: First, for missing data values, missing data for 12 consecutive hours are directly deleted. For other missing data, the K-NearestNeighbours algorithm is used to fill the missing values ​​by averaging the K neighboring samples around the missing samples to maintain the integrity of the data; Secondly, for data outliers, abnormal data that appear repeatedly for 12 consecutive hours are directly deleted, and the Laida criterion is used to filter out outliers that are too large or too small, and the data group containing outliers is deleted.

7. The secondary network heat load prediction method based on CEEMDAN-KPCA-LSTM algorithm according to claim 1 is characterized in that: The input variable selection selects key influencing factors of dynamic changes in heat load of the unit building through the heat transfer process of the unit building and the thermal characteristics of the unit building itself.

8. The secondary network heat load prediction method based on CEEMDAN-KPCA-LSTM algorithm according to claim 1 is characterized in that: The methods for selecting the key influencing factors include, but are not limited to, Pearson correlation analysis, Spearman correlation analysis, Kendall rank, variance analysis, and principal factor analysis.

9. A secondary network heat load prediction device based on CEEMDAN-KPCA-LSTM algorithm, characterized by: include: Data acquisition module: used to obtain historical heating network data and historical environmental factor data of a building during the heating season; Data preprocessing module: used to clean the data set, remove outliers, and use the KNN algorithm to complete the missing data set; Data partitioning module: used to divide the completed data set into training set and test set according to the proportion; Input variable selection module: used to take the key factors that affect the dynamic change of secondary network heat load in the training set data as input variables, and the heat load data as output variables; Data encryption processing module: used for local encryption and standardization of heat load data; Building module: used to build an improved neural network model for predicting dynamic heat loads of buildings; Training module: used to train the improved neural network model. When the improved neural network model meets the convergence conditions, a trained improved neural network model is obtained. Standardization processing module: used to perform standardization processing on the test set data, and set the test set input data according to the input variables selected from the standardized test set data by the training set data; Implementation module: used to input the test set input data into the trained improved neural network model to predict the dynamic heat load of a building.

Images