CNN-LSTM-MHA-based heating load prediction method

Abstract

The invention relates to a heating load prediction method based on CNN-LSTM-MHA, and the method comprises the following steps: carrying out the preprocessing of actual measurement data, and selecting the data of a training set and a test set; establishing a CNN-LSTM-MHA hybrid load prediction model based on a multi-attention mechanism, training the load prediction model, and determining a model architecture; the prediction performance of the model is evaluated, prediction errors are calculated, and the accuracy of the prediction model is evaluated; optimizing the load prediction model by adopting a genetic algorithm GA; and using the optimized load prediction model to predict newly input data and obtain a prediction result. According to the heating load prediction method based on CNN-LSTM-MHA, the internal and external disturbance factors of the building heating load are fully considered as much as possible, on the basis of summarizing the influence of the building thermal response characteristics and the user energy consumption behavior characteristics on the building heating load, an active and passive factor coupled heating load prediction model is constructed, and the prediction precision of the residential building thermal load is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of HVAC feedforward control in the process of building energy management, and in particular to a heating load prediction method based on CNN-LSTM-MHA taking into account the thermal inertia of the building, the influence of active and passive factors, and the difference in human behavior. Background Art

[0002] The heating load is affected by complex nonlinear factors such as the building envelope structure, heating system operation mode, outdoor meteorological conditions, and indoor personnel. The reliability of heating load prediction can be improved by comprehensively considering the fluctuation law of building heating load under the combined effect of various internal and external disturbance factors and adjusting the prediction method based on this.

[0003] According to different prediction methods, existing building load prediction methods can be divided into simulation ("white box"), data-driven ("black box") and hybrid methods ("gray box"). Among them, simulation ("white box") selects the detailed physical properties of existing building materials and uses numerical equations to predict building loads. Currently, the widely used simulation software includes: EnergyPlus, Dest, eQuest and Ecotect. In the calculation process, these methods simplify the behavioral characteristics of personnel; data-driven ("black box") uses a large number of empirical models / statistical models to model thermodynamic problems without specific building parameters. The data-driven method uses representative algorithms, such as regression algorithms or artificial neural networks, to explore the relationship between input and output variables. Hybrid models ("gray box") can improve the problems of poor interpretability of black box models and low accuracy of white box models, but the detailed building information and mathematical assumptions required when constructing hybrid models are still bottlenecks for achieving accurate predictions. In actual situations, the original data often have problems such as missing and noisy data, which will have an adverse effect on the prediction results. In order to improve the reliability of the prediction, data preprocessing is required. In addition, the existing building load prediction models are often unable to comprehensively consider the nonlinear relationship between various internal and external disturbance factors and their impact on the load, such as personnel behavior, building thermal inertia, and the heating characteristics of the pipeline network itself. This can easily lead to a reduction in the reliability of the load prediction model. This is particularly evident in the load prediction of residential buildings. Further data processing is needed to establish a prediction model that considers data from multiple factors to improve the accuracy and reliability of the prediction. Summary of the invention

[0004] The technical problem to be solved by the present invention is: to overcome the defect that the existing white, black and gray box methods for predicting heating load cannot comprehensively consider various internal and external disturbance factors, such as the influence of building thermal inertia, personnel behavior, etc. on the load, resulting in reduced reliability of the load prediction model, and to provide a heating load prediction method based on CNN-LSTM-MHA, which considers the internal and external disturbance factors of the building heating load as fully as possible, and constructs a heating load prediction model of active and passive factors coupling on the basis of summarizing the influence of building thermal response characteristics and user energy consumption behavior characteristics on the building heating load, so as to improve the accurate prediction accuracy of the heat load of residential buildings.

[0005] This heating load forecasting method based on CNN-LSTM-MHA includes the following steps:

[0006] S1-Data processing: Use statistical methods to pre-process the measured data of heating energy consumption of typical residential buildings and select training set and test set data;

[0007] S2-Model establishment and training: Based on the data processed in step S1, a CNN-LSTM-MHA hybrid load forecasting model based on a multi-head attention mechanism is established by combining a convolutional neural network, a long-short time neural network and a multi-head attention mechanism; the load forecasting model is trained using the training set processed in step S1 to verify the adaptability and short-term forecasting ability of the model and determine the model architecture;

[0008] S3-Model evaluation: Use the test set obtained in step S1 to evaluate the prediction performance of the model, calculate the prediction error, analyze the confidence interval of the deviation between the true value and the predicted value, and use the mean absolute error MAE, root mean square error RMSE, absolute percentage error MAPE and determination coefficient R 2 Parameters such as are used to evaluate the accuracy of the prediction model;

[0009] S4-Model optimization: Use genetic algorithm GA to optimize the load forecasting model;

[0010] S5-Load forecasting: Use the load forecasting model optimized in step S4 to forecast the newly input data and obtain the residential building heating load forecast results.

[0011] Furthermore, the data processing process in step S1 includes the following steps:

[0012] S11- Collect heating energy consumption monitoring data, indoor and outdoor temperature and heat flux density data, and investigate the current status of the indoor thermal environment and the indoor behavior characteristics of residents;

[0013] S12-Process the acquired data and information, including data preprocessing and questionnaire information processing, and describe the data using statistical analysis methods and STL methods;

[0014] S13-Divide the data obtained by processing in step S11 into a training set and a test set.

[0015] Furthermore, the load prediction model described in step S2 includes five key layers: input layer, convolutional neural network CNN, long short-term neural network LSTM, attention mechanism Attention Mechanism and fully connected layer; first, in the input layer, a heating energy consumption factor matrix is ​​generated according to the batch size Batch_Size, and sliding processing is performed; then, the input matrix is ​​subjected to a one-dimensional convolution operation in the convolutional neural network CNN through a specific convolution kernel; then, the long short-term neural network LSTM extracts high-level abstract features from the time information dimension output by the convolutional neural network CNN, and maps the input features through the multi-head attention mechanism MHA; finally, the output information is converted into a predicted value of the heating load through the fully connected layer; the key parameters in the model architecture include convolution kernel size, number of LSTM layers, BatchSize, learning rate, exit strategy Droupout, number of heads of the multi-head attention mechanism, optimizer and loss function.

[0016] Furthermore, the method for model optimization in step S4 is to initialize the prediction model structure parameters, train the neural network, calculate the neural network missing, select, cross, and mutate steps in sequence to complete a genetic evolution. When the genetic number Gen meets or exceeds the preset MaxGen, the optimization is completed; the prediction model structure parameters include neural network parameters, population, and genetic number.

[0017] Specifically, the step S3 - model evaluation and the step S4 - model optimization are both performed on the load forecasting model established in the step S2 -.

[0018] The beneficial effect of the heating load prediction method based on CNN-LSTM-MHA of the present invention is as follows: it overcomes the defect that the existing white, black and gray box heating load prediction methods cannot comprehensively consider various internal and external disturbance factors, such as the influence of building thermal inertia, personnel behavior, etc. on the load, resulting in reduced reliability of the load prediction model. It considers the internal and external disturbance factors of the building heating load as much as possible, and constructs a heating load prediction model of active and passive factors coupling on the basis of summarizing the influence of building thermal response characteristics and user energy consumption behavior characteristics on the building heating load, so as to improve the accurate prediction accuracy of the heat load of residential buildings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The following is a further description of a heating load prediction method based on CNN-LSTM-MHA of the present invention in conjunction with the accompanying drawings:

[0020] Figure 1 This is a flowchart of the steps of the heating load forecasting method based on CNN-LSTM-MHA;

[0021] Figure 2 It is a flow chart of the model optimization steps described in the heating load forecasting method based on CNN-LSTM-MHA;

[0022] Figure 3 This is the overall architecture and flow chart of the load forecasting model described in this heating load forecasting method based on CNN-LSTM-MHA. DETAILED DESCRIPTION

[0023] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be directly connected or indirectly connected through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0024] In the description of the present invention, it should be understood that the terms "left", "right", "front", "back", "top", "bottom", "inside", "outside", etc., indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0025] The technical solution of the present invention is further described below with specific embodiments, but the protection scope of the present invention is not limited to the following embodiments.

[0026] Implementation method 1: Figure 1 , 2 As shown, the heating load prediction method based on CNN-LSTM-MHA includes the following steps:

[0027] S1-Data processing: Use statistical methods to pre-process the measured data of heating energy consumption of typical residential buildings and select training set and test set data; the data processing process includes the following steps:

[0028] S11-Use small weather stations, temperature and humidity recorders, heat flow recorders and ultrasonic heat meters to collect heating energy consumption monitoring data, indoor and outdoor temperature and heat flux density data through field tests, and investigate the current status of the indoor thermal environment and the indoor behavior characteristics of residents.

[0029] The specific collection methods and contents are as follows: (1) Outdoor environmental parameters - small meteorological station: temperature, relative humidity, wind speed, solar radiation intensity; (2) Indoor environmental parameters - temperature and humidity recorder: temperature, relative humidity, wind speed; (3) Enclosure structure parameters - heat flow recorder: inner and outer wall temperature and near wall temperature and humidity, inner and outer wall temperature of windows, heat flux density; (4) Heating energy consumption - ultrasonic heat meter: user side supply and return water temperature, cumulative flow, cumulative heat;

[0030] The specific survey method and content are as follows: the test period is from March 4, 2023 to April 24, 2023 (late cold period test) and December 21, 2023 to January 10, 2024 (severe cold period test). Outdoor parameters are recorded at intervals of 30 minutes, and indoor environmental parameters, enclosure structure parameters and heating energy consumption are recorded at intervals of 10 minutes. Through the questionnaire survey task allocation method, the basic situation of households in residential buildings, personal energy consumption behavior, household energy consumption and use of heating equipment, as well as the probability of residents being indoors and indoor comfort are obtained;

[0031] S12-Process the acquired data and information, including data preprocessing and questionnaire information processing, and describe the data using statistical analysis methods and STL methods;

[0032] Data processing method: For numerical data obtained by actual measurement, data preprocessing is required due to the existence of outliers and missing values; for outliers, the box plot method is used to determine that data points that deviate from the upper and lower interquartile range of the box plot by 1.5 times (or other thresholds) are outliers; for missing values ​​of outdoor and indoor environmental parameters, the sliding average method is used to fill the missing values ​​by calculating the average value of continuous windows in the data sequence; for missing values ​​of heating energy consumption data, the heat card filling method is used to fill the missing values ​​by randomly selecting similar samples from existing data; when the missing data accounts for more than 20%, the samples corresponding to the data are deleted;

[0033] Questionnaire survey information processing method: First, the information in the questionnaire needs to be converted and coded according to the different types of questions in the questionnaire to make it data-based; the questions involved in the questionnaire are closed questions, numerical open questions and Likert scale questions. Numerical open questions refer to questions that directly answer numerical values. The answer value in the information processing process is the variable value. Likert scale questions are used to investigate the attitudes of the respondents. The corresponding option values ​​are set according to the attitude tendency scale, and the option values ​​are the variable values.

[0034] The missing values ​​of the information obtained from the questionnaire survey were interpolated and reliability analysis was performed; the missing family member information values ​​were filled with the data of the same type of households, and the missing information values ​​such as personnel energy consumption characteristics and heating satisfaction were filled with the logistic regression method; for Likert scale questions, reliability analysis was required, specifically referring to the analysis of whether a certain type of question in the questionnaire was consistent, which was often evaluated using Cronbach's α coefficient, calculated by formula (1):

[0035]

[0036] In the formula: k – the total number of scale variables;

[0037] ——the variance of the ith variable;

[0038] ——The total variance of all variables.

[0039] S13-dividing the data processed in step S11 into a training set and a test set in a ratio of 8:2 to train and test the heating load prediction model;

[0040] S2-Model establishment and training: Based on the data processed in step S1, a CNN-LSTM-MHA hybrid load forecasting model based on a multi-head attention mechanism is established by combining a convolutional neural network, a long-short time neural network and a multi-head attention mechanism; the load forecasting model is trained using the training set processed in step S1 to verify the adaptability and short-term forecasting ability of the model and determine the model architecture; the model architecture includes the convolution kernel size, the number of LSTM layers, BatchSize, the learning rate, the exit strategy Droupout, the number of heads of the multi-head attention mechanism, the optimizer and the loss function;

[0041] The load forecasting model based on the multi-head attention mechanism adopts a feedforward neural network convolutional neural network (CNN) to extract input parameters (probability of people in the room, time characteristics, indoor and outdoor temperatures; the output data of the prediction model is the heating load) to extract the spatiotemporal characteristics of the input data, which helps the model to better understand the temporal relationship of the input parameters and provide feature mapping for the LSTM layer; the convolutional neural network CNN includes a convolution layer, a pooling layer and a fully connected layer; in the convolution layer, the local features of the data are extracted by convolution operation, and the convolution kernel is used to perform convolution operation on the input. Each convolution kernel generates a set of new feature maps, which are passed to the next layer after being processed by the activation function. The theoretical calculation formula of this process is:

[0042] X n =f(W n *X n-1+B n )(2)

[0043] Where: X n ——The output value of the n-1th layer;

[0044] f——activation function;

[0045] *——convolution symbol;

[0046] W n ——The weight matrix of the convolution kernel input to the nth layer;

[0047] X n-1 ——The output value of the n-1th layer;

[0048] B n ——Bias of the nth layer.

[0049] The pooling layer is used to reduce the dimension of the feature map output by the convolutional layer and aggregate the local area of ​​the input data, which is called maximum pooling. The calculation method is:

[0050]

[0051] Where: w – the kernel size of the pooling layer;

[0052] ——The input value of the tth neuron in the ith feature plane of the lth layer;

[0053] ——The output value of the jth neuron of the ith feature surface in the lth layer.

[0054] The fully connected layer integrates the feature maps extracted by the convolutional layer and the pooling layer into a global feature vector. The convolutional neural network (CNN) gradually learns the mapping relationship between input and output through repeated forward propagation and back propagation, providing more meaningful feature maps for the processing of the long short-term neural network (LSTM).

[0055] The long short-term neural network (LSTM) can selectively memorize and forget the features in the input data, better capture the long-term dependencies in the time series, and improve the reliability of prediction. It specifically includes three structures, namely the forget gate, input gate, and output gate. First, through the forget gate structure, formula (4) is used to calculate the output vector between 0 and 1, which indicates the degree of information retention at the corresponding position.

[0056] f t =σ(W f x t +U f h t-1 +b f )(4)

[0057] Where: f t ——Output vector of the forget gate, value range (0, 1);

[0058] σ——sigmoid activation function;

[0059] W f ——weight matrix;

[0060] h t-1 ——The hidden state of the previous time step;

[0061] x t ——The input for the current time step.

[0062] Then, the input degree of the new information at the corresponding position is obtained through the input gate. Through formula (5), the probability value between 0 and 1 is obtained, which indicates the degree to which the new information at the corresponding position should be added to the memory unit. The closer to 1, the more new information is retained.

[0063] i t =σ(W i x t +U i h t-1 +b i )(5)

[0064] Where: i t ——The output vector of the input gate, the value range is (0, 1);

[0065] σ——sigmoid activation function;

[0066] W i ——weight matrix;

[0067] h t-1 ——The hidden state of the previous time step;

[0068] x t ——The input for the current time step.

[0069] Finally, the output is calculated by the long short-term neural network (LSTM) through the output gate using formula (6);

[0070] o t =σ(W o x t +U o h t-1 +b o )(6)

[0071] The multi-head attention mechanism (MHA) allows the model to dynamically focus on different positions of the input sequence when calculating the output, thereby effectively capturing important information in the input sequence. It solves the limitations of the long short-term neural network (LSTM) in practical applications, such as limited ability to capture long-term dependencies and fixed length memory. In order to enable the prediction model to more flexibly handle input sequences of different lengths and complexities and improve the performance and generalization ability of the prediction model, the multi-head attention mechanism (MHA) is used to merge or splice the outputs of multiple attention heads before processing the fully connected layer of the convolutional neural network (CNN) to obtain the final output;

[0072] The overall prediction process is as follows Figure 3 As shown in the figure, the heating load prediction model contains five key layers: input layer, convolutional neural network (CNN), long short-term neural network (LSTM), attention mechanism (AttentionMechanism) and output layer; in the input layer, the heating energy consumption factor matrix is ​​generated according to the batch size (Batch_Size) and sliding processing is performed; then, the input matrix is ​​subjected to a one-dimensional convolution operation in CNN through a specific convolution kernel; then, LSTM extracts high-level abstract features from the time information dimension of CNN output and maps the input features through a multi-head attention mechanism; finally, the output information is converted into the predicted value of the heating load through the fully connected layer;

[0073] S3-Model evaluation: Use the test set obtained in step S1 to evaluate the prediction performance of the model, calculate the prediction error, analyze the confidence interval of the deviation between the true value and the predicted value, and use the mean absolute error MAE, root mean square error RMSE, absolute percentage error MAPE and determination coefficient R 2 Parameters such as are used to evaluate the accuracy of the prediction model;

[0074] The specific calculation methods of the above indicators are as follows;

[0075]

[0076] Where: ——Predicted value of heating energy consumption at the i-th test point;

[0077] y i ——The actual value of heating energy consumption at the i-th test point;

[0078] n – the number of training sets or test sets.

[0079] At the same time, the test set of supplementary data is used to evaluate the adaptability of the model;

[0080] S4-Model optimization: Genetic algorithm GA is used to optimize the load forecasting model; the model optimization method is to perform initialization of forecasting model structure parameters, neural network training, calculation of neural network missing, selection, crossover, and mutation steps in sequence to complete a genetic evolution, and when the genetic number Gen meets or exceeds the preset MaxGen, the optimization is completed; the forecasting model structure parameters include neural network parameters, population and genetic number;

[0081] like Figure 2 As shown in the figure, during the optimization process, the data set is first loaded to initialize the structural parameters of the prediction model (neural network parameters, population and genetic times); then a neural network with random attributes is generated and trained according to the size of the genetic population, and the neural network loss is calculated. The errors of all models on the training set are sorted, and the best neural network is selected as Xretain. Xretain is the best neural network of the parent generation (g generation) with parent characteristics; after defining the characteristics of the parent generation, the selected parent individuals are combined into a new neural network through a crossover operation. In this stage, only the best performing neural network in the previous generations is selected; in the mutation stage, a certain The probability of randomly changing some attributes of individuals increases the diversity of the search space; the parent generation and individuals obtained through crossover and mutation are merged to form the next generation population (g+1 generation); when the solution with the best fitness is formed, the optimization is considered to be completed and the optimization stops. To determine whether the algorithm stops, it is necessary to check whether the algorithm meets the stopping condition. The stopping condition set in this example is: the number of iterations (Gen) is greater than the set maximum number of iterations (Maxgen), that is, Gen≥Maxgen. When this condition is met, it is considered that a solution with the best fitness is formed and it is used as the output of the algorithm. Otherwise (Gen≤Maxgen), the optimization continues;

[0082] S5-load prediction: Use the load prediction model optimized in step S4 to predict the newly input data to obtain the residential building heating load prediction result;

[0083] After the load forecasting model is optimized, the optimized model can be used for forecasting. The overall process is as shown in the attached figure. Figure 1 As shown in the figure, firstly, the historical energy consumption data of the building is obtained, the outliers and missing values ​​of the data are processed, and the training set and the test set are divided; then, based on the processed data information, a prediction model is constructed by combining convolutional neural network, long-short time neural network and multi-head attention mechanism; a genetic algorithm is used to optimize the hierarchical structure of the prediction model to improve the prediction accuracy; finally, the optimized prediction model is used to carry out prediction. The specific prediction process is shown in the attached figure. Figure 1 As shown, the mean absolute error (MAE), root mean square error (RMSE) and determination coefficient R are selected. 2 Evaluate the accuracy of the prediction model.

[0084] Implementation 2: Step S3-model evaluation and step S4-model optimization of the heating load forecasting method based on CNN-LSTM-MHA are both performed on the load forecasting model established in step S2. The remaining structures and components are as described in Implementation 1 and will not be described again.

[0085] During operation: On the basis of determining the key parameters of the prediction model, the analyzed thermal response characteristics of residential buildings and the energy consumption behavior of personnel are integrated into the heating load prediction process. A CNN-LSTM-MHA hybrid prediction model based on the multi-head attention mechanism is proposed, which is used to predict the heating load of residential buildings, and the accuracy and adaptability of the prediction method are explored. At the same time, the genetic algorithm is used to optimize the structural settings of the prediction model to improve the accuracy of the model prediction and reduce the number of iterations. The heating load prediction model proposed in the study comprehensively considers the influence of indoor thermal response characteristics and personnel energy consumption behavior, can reflect the fluctuation law of heating demand of residential buildings in severe cold areas, accurately predict the changing trend of heating load and maintain stability.

[0086] This heating load forecasting method based on CNN-LSTM-MHA overcomes the defect that the existing white, black and gray box heating load forecasting methods cannot comprehensively consider various internal and external disturbance factors, such as the impact of building thermal inertia and personnel behavior on the load, resulting in reduced reliability of the load forecasting model. It considers the internal and external disturbance factors of the building heating load as much as possible, and constructs a heating load forecasting model coupled with active and passive factors on the basis of summarizing the impact of building thermal response characteristics and user energy consumption behavior characteristics on building heating load, so as to improve the accurate prediction accuracy of residential building heat load.

[0087] The above description shows the main features, basic principles, and advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments or embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, the above embodiments or embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved.

[0088] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A heating load forecasting method based on CNN-LSTM-MHA, characterized by: The Method The following steps are included: S1-Data processing: Use statistical methods to pre-process the measured data of heating energy consumption of typical residential buildings and select training set and test set data; S2-Model establishment and training: Based on the data processed in step S1, a CNN-LSTM-MHA hybrid load forecasting model based on a multi-head attention mechanism is established by combining a convolutional neural network, a long-short time neural network and a multi-head attention mechanism; the load forecasting model is trained using the training set processed in step S1 to verify the adaptability and short-term forecasting ability of the model and determine the model architecture; S3-Model evaluation: Use the test set obtained in step S1 to evaluate the prediction performance of the model, calculate the prediction error, analyze the confidence interval of the deviation between the true value and the predicted value, and use the mean absolute error MAE, root mean square error RMSE, absolute percentage error MAPE and determination coefficient R 2 The accuracy of the prediction model is evaluated by using parameters such as , and the adaptability of the model is evaluated by using a test set of supplementary data; S4-Model optimization: Use genetic algorithm GA to optimize the load forecasting model; S5-Load forecasting: Use the load forecasting model optimized in step S4 to forecast the newly input data and obtain the residential building heating load forecast results.

2. The heating load forecasting method based on CNN-LSTM-MHA according to claim 1 is characterized in that: S1 The data processing process includes the following steps: S11- Collect heating energy consumption monitoring data, indoor and outdoor temperature and heat flux density data, and investigate the current status of the indoor thermal environment and the indoor behavior characteristics of residents; S12-Process the acquired data and information, including data preprocessing and questionnaire information processing, and describe the data using statistical analysis methods and STL methods; S13-Divide the data obtained by processing in step S11 into a training set and a test set.

3. The heating load forecasting method based on CNN-LSTM-MHA according to claim 2 is characterized in that: The load prediction model in step S2 includes five key layers: input layer, convolutional neural network CNN, long short-term neural network LSTM, attention mechanism Attention Mechanism and fully connected layer; first, in the input layer, a heating energy consumption factor matrix is ​​generated according to the batch size Batch_Size and sliding processing is performed; then, the input matrix is ​​subjected to a one-dimensional convolution operation in the convolutional neural network CNN through a specific convolution kernel; then, the long short-term neural network LSTM extracts high-level abstract features from the time information dimension output by the convolutional neural network CNN, and maps the input features through the multi-head attention mechanism MHA; finally, the output information is converted into a predicted value of the heating load through a fully connected layer; the key parameters in the model architecture include convolution kernel size, number of LSTM layers, BatchSize, learning rate, exit strategy Droupout, number of heads of the multi-head attention mechanism, optimizer and loss function.

4. The heating load forecasting method based on CNN-LSTM-MHA according to claim 3 is characterized in that: The method for model optimization in step S4 is to perform initialization of prediction model structure parameters, neural network training, calculation of neural network missing, selection, crossover, and mutation steps in sequence to complete a genetic evolution. When the genetic number Gen meets or exceeds the preset MaxGen, the optimization is completed; the prediction model structure parameters include neural network parameters, population, and genetic number.

5. The heating load forecasting method based on CNN-LSTM-MHA according to claim 4 is characterized in that: The step S3 - model evaluation and the step S4 - model optimization are both performed on the load forecasting model established in the step S2 -.

Images