Public building air conditioning load decomposition analysis method, device and electronic equipment based on two-stage attention mechanism fusion convolutional neural network and long short-term memory network

Abstract

The present invention relates to a method for decomposing air conditioning loads in public buildings, based on a two-stage attention mechanism that integrates a convolutional neural network and a long short-term memory network. The method comprises the following steps: establishing a model for distinguishing between air conditioning operating days and non-operating days based on a clustering algorithm; establishing a non-air conditioning load prediction model using a CNN convolutional neural network combined with an LSTM long short-term memory network; establishing a feature attention mechanism at the input of the non-air conditioning load prediction model to quantify the correlation between input feature variables, dynamically adjust input feature weights, and enhance the guidance of strongly correlated feature variables; establishing a temporal attention mechanism at the output of the non-air conditioning load prediction model to capture the dependencies of temporal information, dynamically adjust the output of hidden states, and enhance the information expression of key time steps; and establishing a public building air conditioning load decomposition model based on a dynamic difference method based on the seasonal operating characteristics of public building loads.

Description

Technical Field

[0001] The present invention discloses a method for decomposing and analyzing the air-conditioning load of public buildings based on a two-stage attention mechanism integrating a convolutional neural network and a long short-term memory network, and relates to a load decomposition method, a device and an electronic device for air-conditioning used in public buildings. Background Art

[0002] Air conditioners, as a device that can regulate indoor temperature, have been widely used. However, as the type of energy consumption with the highest proportion in the whole society, the energy consumption of building air conditioning accounts for about 15% of the total energy consumption of the whole society, and the air conditioning load accounts for more than 30% in summer. In addition, due to the relatively weak awareness of energy conservation in air conditioning, there is a serious problem of energy waste in air conditioning, and there is huge room for energy conservation.

[0003] Current monitoring, analysis, and interactive control of air conditioning loads are plagued by a lack of comprehensive building air conditioning monitoring data, relatively crude air conditioning load estimation methods, and a lack of benchmarking methods for air conditioning energy consumption. In particular, for public buildings without air conditioning sub-item measurement data, the power grid can only estimate user air conditioning load by subtracting the real-time threshold load from a typical daily baseline load. This method suffers from poor real-time performance and accuracy. Therefore, constructing a neural network model to decompose user air conditioning loads in real time is of great practical significance for the power grid to accurately assess user energy consumption and formulate appropriate energy-saving strategies. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a decomposition and analysis method for the air-conditioning load of public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory (dual attention mechanism-convolutional neural network-longshort-term memory, DA-CNN-LSTM) network. A public building air-conditioning load decomposition method based on a CNN-LSTM network is established, and the accuracy of the model is improved by establishing a dual attention mechanism (DA), providing a powerful analysis tool for studying the energy consumption and adjustable potential of air-conditioning in public buildings.

[0005] The present invention provides a decomposition and analysis method for air conditioning load in public buildings based on a two-stage attention mechanism integrating a convolutional neural network and a long short-term memory network, comprising the following steps:

[0006] Based on the clustering algorithm, a model was established to distinguish between air conditioning operation days and non-operation days. The air conditioning non-operation days after differentiation were used as the training set of the neural network prediction model, and the air conditioning operation days were used as the test set of the model.

[0007] A non-air conditioning load forecasting model is established by combining CNN convolutional neural network with LSTM long short-term memory network;

[0008] A feature attention mechanism is established on the input side of the non-air-conditioning load forecasting model to quantify the correlation between input feature variables, thereby dynamically adjusting the input feature weights and enhancing the guidance of strongly correlated feature variables.

[0009] A temporal attention mechanism is established on the output side of the non-air-conditioning load forecasting model to capture the dependencies of temporal information, dynamically adjust the output of hidden states, and enhance the information expression of key time steps.

[0010] Based on the results of distinguishing between air-conditioning operation days and non-operation days in public buildings, a public building air-conditioning load decomposition method based on dynamic subtraction method is established, that is, the air-conditioning load of public buildings is obtained by dynamically subtracting the threshold load curve of air-conditioning operation days from the non-air-conditioning load curve obtained by the non-air-conditioning load prediction model.

[0011] Furthermore, the method for establishing a model for distinguishing between air-conditioning operation days and non-operation days based on the clustering algorithm is as follows:

[0012] Based on the clustering algorithm, a clustering method for the threshold load curves of public building users was established to distinguish between days when the user's air conditioner is turned on and days when it is not in operation. The user's daily threshold load curve was input into the clustering algorithm and clustered into two clusters.

[0013] Select a clustering algorithm and use the actual sample that is most similar to other samples in each cluster as the cluster center to improve the clustering algorithm's sensitivity to noise and outliers and its tendency to fall into local optimality;

[0014] Using a clustering algorithm, the output is a collection of two different curves, each representing an air conditioner-on day and an air conditioner-off day. This method is more accurate than using empirical evidence, such as seasonal temperature fluctuations, to distinguish between air conditioner-on and -off days.

[0015] Furthermore, the method of selecting the clustering algorithm is:

[0016] First, for a given cluster number of 2, two representative objects are randomly selected as the initial cluster centers, the distance between each remaining object and the representative object is calculated and assigned to the nearest cluster, and the corresponding clustering results are generated;

[0017] Then, the iterative process begins. For each iteration, a randomly selected non-center point replaces one of the original center points and the clustering results are recalculated;

[0018] If the clustering effect is improved, the replacement is retained, otherwise the original center point is restored. When the replacement no longer improves the clustering effect, the iteration stops;

[0019] The cost function is used to measure the quality of the clustering results. This function is used to measure the average dissimilarity between the object and the center point. The specific definition is as follows:

[0020]

[0021] Among them, p is a point in space, that is, a given object; o i Cluster C i The center point of the dataset; E is the sum of squared deviations of all objects in the dataset.

[0022] Furthermore, the method of establishing a non-air-conditioning load prediction model by using a CNN convolutional neural network combined with an LSTM long short-term memory network is as follows:

[0023] Establish the factors affecting the load of public buildings, including historical load, meteorological information and date type;

[0024] The historical load includes the historical load of the past day, the load at the same time in the past week, and the load at the same time on the same characteristic day in the past month.

[0025] Meteorological information includes temperature, humidity, wind speed and weather;

[0026] Date types include time variables, weekday and rest day variables, and holiday variables.

[0027] The load influencing factors of public building users are used as the input features of the training set and input into the CNN-LSTM network model through a sliding window. The non-air-conditioning load of public building users in the non-air-conditioning season is used as the target feature of the training set to train the CNN-LSTM network.

[0028] Establish CNN-LSTM network model training samples;

[0029] Among them, in the CNN convolutional neural network, the input layer, convolution layer, pooling layer, convolution layer, pooling layer, fully connected layer and output layer are connected in sequence.

[0030] Using the trained model, the load influencing factors at the corresponding time of the air-conditioning season and the non-air-conditioning load of a typical day in the non-air-conditioning season are input from the input layer in the form of a sliding window. The trained model parameters are obtained through the CNN layer and the three-layer LSTM layer. The predicted non-air-conditioning load values ​​at each time of each day in the air-conditioning season are output through the output layer.

[0031] Furthermore, the method for establishing the CNN-LSTM network model training sample is:

[0032] The CNN network significantly reduces the number of parameters in the training process through local neuron connections and convolution kernel weight sharing in the convolution layer operation, improves the model training speed, and enables the model to more effectively extract feature information from the original data;

[0033] In the pooling layer, the original data is abstracted and understood to reduce the feature dimension, which effectively reduces the number of training parameters while reducing the degree of model overfitting and improving the efficiency of feature data extraction;

[0034] The LSTM (Long Short-Term Memory) network evolved from the RNN (Recurrent Neural Network). LSTM has an LSTM memory block. Based on the traditional RNN network structure of input, hidden, and output layers, it improves the hidden layer structure. It introduces a gate mechanism to control the path of information transmission and selectively memorizes or deletes information passing through the network through gates. This allows for selective memorization of valid information and deletion of invalid, redundant information during long-term predictions.

[0035] The state of the LSTM memory block in the corresponding time period is the cell state, which is the old cell state, the current cell state, and the updated cell state in chronological order;

[0036] Each memory block has three "gating" structures, including input gate, output gate and forget gate, through which the cell state is protected and controlled, that is, time information is stored and transmitted;

[0037] The forget gate is responsible for deciding how much of the cell state at the previous moment to retain in the cell state at the current moment, that is, deciding what information to discard from the cell state. The formula is:

[0038] f t =σ(W f *[h t-1 ,x t ]+b f )

[0039] Where h t-1 and h t are the hidden layer states at time t-1 and t, respectively, x t-1 and x t are the input values ​​at time t-1 and t respectively, W f is the weight matrix, b f is the bias, σ is the sigmoid function;

[0040] The input gate is responsible for deciding how much of the current input to the current cell state to retain. The formula is:

[0041] i t =σ(W i *[h t-1 ,x t ]+b i )

[0042]

[0043] Where i tis the input value of the input gate at time t, Create a new cell state value vector for the tanh layer at time t, W i and W C is the weight coefficient matrix, b i and b C is the bias term;

[0044] Update the old cell state, the formula is:

[0045]

[0046] Where C t-1 and C t are the cell states at time t-1 and t respectively;

[0047] The output gate is responsible for determining how many outputs the current cell state has. The formula is:

[0048] o t =σ(W o *[h t-1 ,x t ]+b o )

[0049] h t =o t *tanh(C t )

[0050] In the formula, o t is the output value of the output gate at time t, W o is the weight coefficient matrix, b o is the bias term.

[0051] Furthermore, the method of establishing a feature attention mechanism on the input side of the non-air-conditioning load forecasting model, quantifying the correlation between input feature variables, dynamically adjusting the input feature weights, and enhancing the guidance of strongly correlated feature variables is as follows;

[0052] On the input side, a feature attention mechanism is established to explore the correlation between the input features of weather, time, and historical load and the non-air conditioning load value at the corresponding moment. The attention weight of each input feature is dynamically allocated to focus on the impact of important feature variables and reduce the information of weakly related or redundant features.

[0053] In order to obtain the correlation between each influencing feature and the current non-air-conditioning load, the feature attention mechanism is used to encode the quantized weights;

[0054] After establishing the feature attention mechanism, the CNN-LSTM network no longer inputs raw feature values, but instead uses weighted input feature values ​​that take into account the degree of correlation. This feature attention mechanism adaptively extracts the contribution of each input feature to improve prediction accuracy.

[0055] Furthermore, the method of encoding quantized weights using the feature attention mechanism is:

[0056] The input features at the current moment and the hidden layer state h at the previous moment t-1 and memory information t-1 As the input of the feature attention mechanism, the attention weight of each input feature at the current moment is calculated respectively, and quantized using the calculation method of the multi-layer perceptron. The weight calculation is as follows:

[0057]

[0058] Where V e 、W e and U e is the multi-layer perceptron weight matrix for calculating attention weights, b e is the bias term;

[0059] Use the Softmax function to Normalize the feature attention weights so that the sum is 1, as shown in the following formula:

[0060]

[0061] The characteristic correlation coefficient The corresponding input affects the eigenvalue Multiply to get the associated features that consider the contribution rate of different inputs to the features

[0062]

[0063] Hidden state h t Then update as follows:

[0064]

[0065] Among them, f CNN-LSTM is the LSTM network unit.

[0066] Furthermore, the method of establishing a temporal attention mechanism on the output side of the non-air-conditioning load forecasting model, capturing the dependency of temporal information, dynamically adjusting the output of hidden states, and enhancing the information expression of key time steps is as follows;

[0067] A temporal attention mechanism is established on the decoding output side to obtain the hidden state weights of the corresponding time series, and to determine the degree of influence of the past time series state on the current non-air-conditioning load output, thereby strengthening the influence of strongly correlated time state information.

[0068] In order to obtain the correlation between the historical time series information and the current output non-air conditioning load value, the temporal attention mechanism is used to decode the quantized weights;

[0069] Establishing a dual attention mechanism

[0070] Based on the existing CNN-LSTM network, a feature attention mechanism is introduced before the first LSTM layer, and a time attention mechanism is introduced after the last LSTM layer. This dual attention mechanism can adaptively mine the relationship between load and influencing factors, obtain the contribution rate of each influencing factor at different times, and improve the accuracy of single-point predictions; it can also mine the correlation between the current load and historical key moments, and improve the accuracy of long-term time series predictions.

[0071] That is, the threshold load and load influencing factors in the non-air-conditioning season pass through the input layer, convolution layer, pooling layer, feature attention mechanism, LSTM layer 1, LSTM layer 2, LSTM layer 3, temporal attention mechanism, fully connected layer and output layer in sequence to obtain the non-air-conditioning load in the air-conditioning season; LSTM layer 1, LSTM layer 2 and LSTM layer 3 are LSTM layers, and the convolution layer, pooling layer and fully connected layer are CNN layers.

[0072] Furthermore, the method of decoding quantized weights of the temporal attention mechanism is:

[0073] The hidden layer state h containing the time series information at each historical moment τ (τ=1,2,…,t,…,T) and the hidden layer state h at the previous moment of historical moment t t and memory information t As the input of the temporal attention mechanism, the attention weights of the historical moment t and each historical moment are calculated respectively, and quantified using the calculation method of the multi-layer perceptron. The weight calculation is as follows:

[0074]

[0075] In the formula, V d and W d is the multi-layer perceptron weight matrix for calculating attention weights, U d is the bias term;

[0076] Use the Softmax function to Normalize the weights so that the sum of the temporal attention weights is 1, as shown in the following formula:

[0077]

[0078] Quantify the importance of the state information of the τth hidden layer to the predicted output at time t, and set all the weights The weighted summation of the corresponding hidden layer state information is used to obtain the comprehensive information c related to the temporal state characteristics. t , as shown below:

[0079]

[0080] The non-air-conditioning load value y to be predicted at time t t Comprehensive information related to temporal state characteristics c t After fusion through a single-layer feedforward network, the input of the CNN-LSTM network at each time t is obtained

[0081]

[0082] Where, and The weights and biases of a single-layer feedforward network are calculated to fusion input respectively;

[0083] h t The CNN-LSTM network in the following formula iterates T times from the initial moment to the T moment:

[0084]

[0085] The hidden layer state h at the final moment T T and comprehensive information T The single-layer feedforward network is used for integration and finally the dimension transformation is performed by the single-layer feedforward network to obtain the predicted value of the non-air-conditioning load to be predicted. As shown in the following formula:

[0086]

[0087] Where W y and b w are the weights and biases of the single-layer feedforward network that integrates the output, V y and b y are the weights and biases of the feedforward network for dimension transformation, respectively.

[0088] Furthermore, the modeling method for establishing a public building air conditioning load decomposition model based on the dynamic difference method based on the seasonal operating characteristics of the public building load is as follows:

[0089] The building is divided into air-conditioning season and non-air-conditioning season according to whether the air conditioner is turned on or not, and the following constraints are established:

[0090] P 非空调季关口负荷 ≈P 非空调季非空调负荷

[0091] P 空调季关口负荷 =P 空调季非空调负荷 +P 空调季空调负荷

[0092] Establish a method to dynamically predict the non-air-conditioning load in the air-conditioning season by using the threshold load in the non-air-conditioning season;

[0093] Due to the following constraints:

[0094] P 空调负荷 =P 关口负荷 -P 非空调负荷

[0095] For the air-conditioning operation days divided by clustering, a public building air-conditioning load decomposition method is established, which uses the real-time measured gateway load actual value and the non-air-conditioning load prediction value dynamically predicted by the neural network to decompose the dynamic prediction value of the air-conditioning load. Based on the non-air-conditioning load prediction result of the air-conditioning season, the air-conditioning load prediction value of the air-conditioning season is obtained by real-time difference.

[0096] An electronic device includes one or more processors for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors perform a decomposition and analysis method for air conditioning loads for public buildings based on a DA-CNN-LSTM network.

[0097] A storage medium stores a computer program, wherein the computer program is configured to execute a decomposition and analysis method for air conditioning loads for public buildings using a DA-CNN-LSTM network when running.

[0098] The beneficial effects of the present invention are:

[0099] (1) A non-air conditioning load forecasting method was established based on the CNN-LSTM neural network;

[0100] (2) The accuracy of the load forecasting model is improved by introducing the dual attention mechanism (DA);

[0101] (3) Based on the seasonal characteristics of air-conditioning load in public buildings, a method was established to predict the non-air-conditioning load in the air-conditioning season by using the non-air-conditioning threshold load. The air-conditioning load of public buildings was then decomposed by differentiating the measurable real-time threshold load of the air-conditioning season with the predicted non-air-conditioning load. This method provides a powerful analytical tool for studying the energy consumption and adjustable potential of air-conditioning in public buildings.

[0102] Description of the attached figures and tables

[0103] Figure 1 Schematic diagram of the method of the present invention.

[0104] Figure 2 This is a diagram of the CNN network structure used in the present invention.

[0105] Figure 3 This is a diagram of the LSTM network structure used in the present invention.

[0106] Figure 4 This is a structural diagram of the CNN-LSTM network model used in the present invention.

[0107] Figure 5 Schematic diagram of the feature attention mechanism (FA) used in this invention.

[0108] Figure 6 Schematic diagram of the temporal attention mechanism (TA) used in this invention.

[0109] Figure 7 This is a structural diagram of the DA-CNN-LSTM network model used in the present invention.

[0110] Figure 8 This is a structural diagram of the public building air conditioning load decomposition method used in the present invention.

[0111] Figure 9 This is a structural schematic diagram of the analysis device of the public building air-conditioning load decomposition analysis method based on the two-stage attention mechanism fusing convolutional neural network and long short-term memory network in the present invention.

[0112] Figure 10 This is a schematic diagram of the structure of a computer device in some embodiments of the present invention. DETAILED DESCRIPTION

[0113] In order to more clearly illustrate the technical solution of the present invention, the present invention will be further elaborated below. Obviously, the following description is only a part of the embodiments. For ordinary technicians in this field, they can also apply the technical solution of the present invention to other similar scenarios based on these embodiments without paying any creative work. In order to more clearly illustrate the technical solution of the present invention, the following description is combined with the attached Figure 1-10 The technical solution of the present invention is described in further detail.

[0114] Example 1:

[0115] like Figure 1-8 As shown in FIG, a decomposition and analysis method for air conditioning load in public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network includes the following steps:

[0116] The present invention provides a decomposition and analysis method for air conditioning load in public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network, comprising the following steps:

[0117] Based on a clustering algorithm (K-Medoids), a model was established to distinguish between air conditioning operating days and non-operating days. The non-operating days after differentiation were used as the training set of the neural network prediction model, and the operating days were used as the test set of the model.

[0118] A non-air conditioning load forecasting model is established by combining CNN convolutional neural network with LSTM long short-term memory network;

[0119] A feature attention mechanism (FA) is established on the input side of the non-air-conditioning load forecasting model to quantify the correlation between input feature variables, thereby dynamically adjusting the input feature weights and enhancing the guidance of strongly correlated feature variables.

[0120] A temporal attention mechanism (TA) is established on the output side of the non-air-conditioning load forecasting model to capture the dependencies of temporal information, dynamically adjust the output of hidden states, and enhance the information expression of key time steps.

[0121] Based on the results of distinguishing between air-conditioning operation days and non-operation days in public buildings, a public building air-conditioning load decomposition method based on dynamic subtraction method is established, that is, the air-conditioning load of public buildings is obtained by dynamically subtracting the threshold load curve of air-conditioning operation days from the non-air-conditioning load curve obtained by the non-air-conditioning load prediction model.

[0122] Example 2:

[0123] The present invention provides a modeling method for establishing a model for distinguishing between air-conditioning operation days and non-operation days, and the modeling method is as follows:

[0124] Based on the clustering algorithm (K-Medoids), a clustering method for the threshold load curves of public building users was established to distinguish between days when the user's air conditioner was turned on and days when the air conditioner was not turned on. The user's daily threshold load curves were input into the clustering algorithm and clustered into two clusters.

[0125] A clustering algorithm (K-Medoids) is selected, and the actual sample that is most similar to other samples in each cluster is used as the cluster center to improve the defect that the K-Means clustering algorithm is sensitive to noise and outliers and is prone to falling into local optimality.

[0126] The specific steps of the clustering algorithm are:

[0127] First, for a given cluster number of 2, two representative objects are randomly selected as the initial cluster centers, the distance between each remaining object and the representative object is calculated and assigned to the nearest cluster, and the corresponding clustering results are generated;

[0128] Then, the iterative process begins. For each iteration, a randomly selected non-center point replaces one of the original center points and the clustering results are recalculated;

[0129] If the clustering effect is improved, the replacement is retained, otherwise the original center point is restored. When the replacement no longer improves the clustering effect, the iteration stops.

[0130] The cost function is used to measure the quality of the clustering results. This function is used to measure the average dissimilarity between the object and the center point. The specific definition is as follows:

[0131]

[0132] Among them, p is a point in space, that is, a given object; o i Cluster C i The center point of the dataset; E is the sum of squared deviations of all objects in the dataset.

[0133] Using a clustering algorithm (K-Medoids), the output is a cluster of two different curves corresponding to the dates of the air conditioner operation and non-operation days. This method is more accurate than using empirical evidence such as seasonal temperature fluctuations to distinguish between air conditioner on and off days.

[0134] Example 3:

[0135] The method for establishing the non-air conditioning load prediction model in the present invention is as follows:

[0136] Establish the factors affecting the load of public buildings, including historical load, meteorological information and date type;

[0137] The historical load includes the historical load of the past day, the load at the same time in the past week, and the load at the same time on the same characteristic day in the past month.

[0138] Meteorological information includes temperature, humidity, wind speed and weather;

[0139] Date types include time variables, weekday and rest day variables, and holiday variables.

[0140] The load influencing factors of public building users are used as the input features of the training set and input into the CNN-LSTM network model through a sliding window. The non-air-conditioning load of public building users in the non-air-conditioning season is used as the target feature of the training set to train the CNN-LSTM network.

[0141] Establish CNN-LSTM network model training samples;

[0142] Refer to the attached Figure 2In the CNN convolutional neural network, the input layer, convolution layer, pooling layer, convolution layer, pooling layer, fully connected layer and output layer are connected in sequence.

[0143] In the convolutional layer operation, the CNN network greatly reduces the number of parameters in the training process through local connections of neurons and sharing of convolution kernel weights, improves the model training speed, and enables the model to more effectively extract feature information from the original data.

[0144] In the pooling layer, the original data is abstracted and understood, the feature dimension is reduced, which effectively reduces the number of training parameters while reducing the degree of model overfitting and improving the efficiency of feature data extraction.

[0145] The LSTM long short-term memory network evolved from the RNN recurrent neural network. LSTM has an LSTM memory block. Based on the structure of the traditional RNN network input layer, hidden layer, and output layer, it improves the structure of the hidden layer. By introducing a gate mechanism to control the path of information transmission, and selectively memorizing or deleting information passing through the network through the gate, LSTM selectively memorizes valid information and deletes invalid redundant information during long-term predictions.

[0146] like Figure 3 The LSTM memory block shown, the state of the LSTM memory block in the corresponding time period is the cell state, which is the old cell state, the current cell state, and the updated cell state in chronological order;

[0147] Each memory block has three "gating" structures, including input gate, output gate and forget gate, which are used to protect and control the cell state, that is, to store and transmit time information.

[0148] The forget gate is responsible for deciding how much of the cell state at the previous moment to retain in the cell state at the current moment, that is, deciding what information to discard from the cell state. The formula is:

[0149] f t =σ(W f *[h t-1 ,x t ]+b f )

[0150] Where h t-1 and h t are the hidden layer states at time t-1 and t, respectively, x t-1 and x t are the input values ​​at time t-1 and t respectively, W f is the weight matrix, b f is the bias, σ is the sigmoid function;

[0151] The input gate is responsible for deciding how much of the current input to the current cell state to retain. The formula is:

[0152] i t =σ(W i *[h t-1 ,x t ]+b i )

[0153]

[0154] Where i t is the input value of the input gate at time t, Create a new cell state value vector for the tanh layer at time t, W i and W C is the weight coefficient matrix, b i and b C is the bias term;

[0155] Update the old cell state, the formula is:

[0156]

[0157] Where C t-1 and C t are the cell states at time t-1 and t respectively;

[0158] The output gate is responsible for determining how many outputs the current cell state has. The formula is:

[0159] o t =σ(W o *[h t-1 ,x t ]+b o )

[0160] h t =o t *tanh(C t )

[0161] In the formula, o t is the output value of the output gate at time t, W o is the weight coefficient matrix, b o is the bias term.

[0162] Use a trained model such as Figure 4 As shown in the figure, the load influencing factors at the corresponding time of the air-conditioning season and the non-air-conditioning load of a typical day in the non-air-conditioning season are input from the input layer in the form of a sliding window. The trained model parameters are obtained through the CNN layer and the three-layer LSTM layer. The predicted non-air-conditioning load values ​​at each time of each day in the air-conditioning season are output through the output layer.

[0163] Specifically, the threshold load and load influencing factors in the non-air-conditioning season pass through the input layer, convolution layer, pooling layer, LSTM layer 1, LSTM layer 2, LSTM layer 3, fully connected layer and output layer in sequence to obtain the non-air-conditioning load in the air-conditioning season; LSTM layer 1, LSTM layer 2 and LSTM layer 3 are LSTM layers, and the convolution layer, pooling layer and fully connected layer are CNN layers.

[0164] Example 4:

[0165] The method of establishing a time attention mechanism on the output side of the non-air-conditioning load forecasting model in the present invention is as follows:

[0166] On the input side, a feature attention mechanism (FA) is established to explore the correlation between the input features of weather, time and historical load and the non-air-conditioning load value at the corresponding moment, dynamically allocate the attention weight of each input feature, focus on the influence of important feature variables, and reduce the information of weakly correlated or redundant irrelevant features.

[0167] In order to obtain the correlation between each influencing feature and the non-air-conditioning load at the current moment, the feature attention mechanism is used to encode the quantized weights.

[0168] Refer to the attached Figure 5 , the input features at the current moment and the hidden layer state h at the previous moment t-1 and memory information t-1 As the input of the feature attention mechanism, the attention weight of each input feature at the current moment is calculated respectively, and quantized using the calculation method of the multi-layer perceptron. The weight calculation is as follows:

[0169]

[0170] Where V e 、W e and U e is the multi-layer perceptron weight matrix for calculating attention weights, b e is the bias term.

[0171] Use the Softmax function to Normalize the feature attention weights so that the sum is 1, as shown in the following formula:

[0172]

[0173] The characteristic correlation coefficient The corresponding input affects the eigenvalue Multiply to get the associated features that consider the contribution rate of different inputs to the features

[0174]

[0175] Hidden state h t Then update as follows:

[0176]

[0177] Among them, f CNN-LSTM is the LSTM network unit.

[0178] With the Feature Attention (FA) mechanism, the CNN-LSTM network no longer uses raw input feature values ​​as input, but instead uses weighted input feature values ​​that take into account the degree of correlation. This feature attention mechanism adaptively extracts the contribution of each input feature to improve prediction accuracy.

[0179] Example 5:

[0180] The method of establishing a time attention mechanism on the output side of the non-air-conditioning load forecasting model in the present invention is as follows:

[0181] A temporal attention mechanism (TA) is established on the decoding output side to obtain the hidden state weights of the corresponding time series, and the influence of the time series state at the past moment on the non-air-conditioning load output at the current moment is obtained, thereby strengthening the influence of the strongly correlated moment state information.

[0182] In order to obtain the correlation between the historical time series information and the non-air-conditioning load value output at the current moment, the temporal attention mechanism is used to decode the quantization weights.

[0183] Refer to the attached Figure 6 , the hidden layer state h containing time series information at each historical moment τ (τ=1,2,…,t,…,T) and the hidden layer state h at the previous moment of historical moment t t and memory information t As the input of the temporal attention mechanism, the attention weights of the historical moment t and each historical moment are calculated respectively, and quantified using the calculation method of the multi-layer perceptron. The weight calculation is as follows:

[0184]

[0185] In the formula, V d and W d is the multi-layer perceptron weight matrix for calculating attention weights, U d is the bias term.

[0186] Use the Softmax function to Normalize the weights so that the sum of the temporal attention weights is 1, as shown in the following formula:

[0187]

[0188] Quantify the importance of the state information of the τth hidden layer to the predicted output at time t, and set all the weights The weighted summation of the corresponding hidden layer state information is used to obtain the comprehensive information c related to the temporal state characteristics. t , as shown below:

[0189]

[0190] The non-air-conditioning load value y to be predicted at time t t Comprehensive information related to temporal state characteristics c t After fusion through a single-layer feedforward network, the input of the CNN-LSTM network at each time t is obtained

[0191]

[0192] Where, and The weights and biases of a single-layer feed-forward network are calculated for the fused input.

[0193] h t The CNN-LSTM network in the following formula iterates T times from the initial moment to the T moment:

[0194]

[0195] The hidden layer state h at the final moment T T and comprehensive information T The single-layer feedforward network is used for integration and finally the dimension transformation is performed by the single-layer feedforward network to obtain the predicted value of the non-air-conditioning load to be predicted. As shown in the following formula:

[0196]

[0197] Where W y and b w are the weights and biases of the single-layer feedforward network that integrates the output, V y and b y are the weights and biases of the feedforward network for dimension transformation, respectively.

[0198] Establishing a dual attention mechanism (DA)

[0199] Refer to the attached Figure 7Based on the existing CNN-LSTM network, a feature attention mechanism (FA) is introduced before the first LSTM layer, and a temporal attention mechanism (TA) is introduced after the last LSTM layer. This dual attention mechanism (DA) can adaptively explore the relationship between load and influencing factors, determining the contribution of each influencing factor at different times, improving the accuracy of single-point predictions. It can also explore the correlation between current load and historical key moments, improving the accuracy of long-term time series predictions.

[0200] That is, the threshold load and load influencing factors in the non-air-conditioning season pass through the input layer, convolution layer, pooling layer, feature attention mechanism, LSTM layer 1, LSTM layer 2, LSTM layer 3, temporal attention mechanism, fully connected layer and output layer in sequence to obtain the non-air-conditioning load in the air-conditioning season; LSTM layer 1, LSTM layer 2 and LSTM layer 3 are LSTM layers, and the convolution layer, pooling layer and fully connected layer are CNN layers.

[0201] Example 6:

[0202] The method of establishing the public building air conditioning load decomposition method based on the dynamic difference method is as follows:

[0203] The building is divided into air-conditioning season and non-air-conditioning season according to whether the air conditioner is turned on or not, and the following constraints are established:

[0204] P 非空调季关口负荷 ≈P 非空调季非空调负荷

[0205] P 空调季关口负荷 =P 空调季非空调负荷 +P 空调季空调负荷

[0206] Establish a method to dynamically predict the non-air-conditioning load in the air-conditioning season through the threshold load in the non-air-conditioning season.

[0207] Due to the following constraints:

[0208] P 空调负荷 =P 关口负荷 -P 非空调负荷

[0209] For the air-conditioning operation days identified by clustering, a public building air-conditioning load decomposition method is established, which uses the real-time measurement of the actual value of the gateway load and the non-air-conditioning load prediction value dynamically predicted by the neural network to decompose the dynamic prediction value of the air-conditioning load. The complete process is as follows: Figure 8 As shown, based on the non-air-conditioning load forecast results for the air-conditioning season, the air-conditioning load forecast value for the air-conditioning season is obtained by real-time difference.

[0210] Example 7:

[0211] According to an embodiment of the present invention, an electronic device is also provided, comprising one or more processors for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors execute the aforementioned decomposition and analysis method for air conditioning load for public buildings based on the DA-CNN-LSTM network.

[0212] Example 8:

[0213] According to an embodiment of the present invention, a storage medium is further provided, in which a computer program is stored, wherein the computer program is configured to execute the aforementioned decomposition and analysis method for air-conditioning load for public buildings based on the DA-CNN-LSTM network when running.

Claims

1. A decomposition and analysis method for air conditioning load in public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network, characterized in that: The steps include: Based on the clustering algorithm, a model was established to distinguish between air conditioning operation days and non-operation days. The air conditioning non-operation days after differentiation were used as the training set of the neural network prediction model, and the air conditioning operation days were used as the test set of the model. A non-air conditioning load forecasting model is established by combining CNN convolutional neural network with LSTM long short-term memory network; A feature attention mechanism is established on the input side of the non-air-conditioning load forecasting model to quantify the correlation between input feature variables, thereby dynamically adjusting the input feature weights and enhancing the guidance of strongly correlated feature variables. A temporal attention mechanism is established on the output side of the non-air-conditioning load forecasting model to capture the dependencies of temporal information, dynamically adjust the output of hidden states, and enhance the information expression of key time steps. Based on the results of distinguishing between air-conditioning operation days and non-operation days in public buildings, a method for decomposing the air-conditioning load of public buildings based on the dynamic subtraction method is established. That is, the air-conditioning load of public buildings is obtained by dynamically subtracting the threshold load curve of the air-conditioning operation day from the non-air-conditioning load curve obtained by the non-air-conditioning load prediction model. The method for establishing a model for distinguishing between air-conditioning operation days and non-operation days based on the clustering algorithm is as follows: Based on the clustering algorithm, a clustering method for the threshold load curves of public building users was established to distinguish between days when the user's air conditioner is turned on and days when it is not in operation. The user's daily threshold load curve was input into the clustering algorithm and clustered into two clusters. Select a clustering algorithm and use the actual sample that is most similar to other samples in each cluster as the cluster center to improve the clustering algorithm's sensitivity to noise and outliers and its tendency to fall into local optimality; Using a clustering algorithm, the output of two different curves after classification corresponds to the dates representing the air conditioner operating days and non-operating days. This method is more accurate than distinguishing between air conditioner on / off days based on experience such as seasonal temperature changes. The method of selecting the clustering algorithm is: First, for a given cluster number of 2, two representative objects are randomly selected as the initial cluster centers, the distance between each remaining object and the representative object is calculated and assigned to the nearest cluster, and the corresponding clustering results are generated; Then, the iterative process begins. For each iteration, a randomly selected non-center point replaces one of the original center points and the clustering results are recalculated; If the clustering effect is improved, keep the replacement, otherwise restore the original center point; When the replacement no longer improves the clustering effect, the iteration stops; The cost function is used to measure the quality of the clustering results. This function is used to measure the average dissimilarity between the object and the center point. The specific definition is as follows: Among them, p is a point in space, that is, a given object; o i Cluster C i The center point of the dataset; E is the sum of squared deviations of all objects in the dataset.

2. The method for decomposing and analyzing the air conditioning load for public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network according to claim 1 is characterized in that: The method for establishing a non-air conditioning load prediction model by using a CNN convolutional neural network combined with an LSTM long short-term memory network is as follows: Establish the factors affecting the load of public buildings, including historical load, meteorological information and date type; The historical load includes the historical load of the past day, the load at the same time in the past week, and the load at the same time on the same characteristic day in the past month. Meteorological information includes temperature, humidity, wind speed and weather; Date types include time variables, weekday and rest day variables, and holiday variables; The load influencing factors of public building users are used as the input features of the training set and input into the CNN-LSTM network model through a sliding window. The non-air-conditioning load of public building users in the non-air-conditioning season is used as the target feature of the training set to train the CNN-LSTM network. Establish CNN-LSTM network model training samples; Among them, in the CNN convolutional neural network, the input layer, convolution layer, pooling layer, convolution layer, pooling layer, fully connected layer and output layer are connected in sequence; Using the trained model, the load influencing factors at the corresponding time of the air-conditioning season and the non-air-conditioning load of a typical day in the non-air-conditioning season are input from the input layer in the form of a sliding window. The trained model parameters are obtained through the CNN layer and the three-layer LSTM layer. The predicted non-air-conditioning load values ​​at each time of each day in the air-conditioning season are output through the output layer.

3. The method for decomposing and analyzing the air conditioning load for public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network according to claim 2 is characterized in that: The method for establishing the CNN-LSTM network model training sample is: The CNN network significantly reduces the number of parameters in the training process through local neuron connections and convolution kernel weight sharing in the convolution layer operation, improves the model training speed, and enables the model to more effectively extract feature information from the original data; In the pooling layer, the original data is abstracted and understood to reduce the feature dimension, which effectively reduces the number of training parameters while reducing the degree of model overfitting and improving the efficiency of feature data extraction; The LSTM (Long Short-Term Memory) network evolved from the RNN (Recurrent Neural Network). LSTM has an LSTM memory block. Based on the traditional RNN network structure of input, hidden, and output layers, it improves the hidden layer structure. It introduces a gate mechanism to control the path of information transmission and selectively memorizes or deletes information passing through the network through gates. This allows for selective memorization of valid information and deletion of invalid, redundant information during long-term predictions. The state of the LSTM memory block in the corresponding time period is the cell state, which is the old cell state, the current cell state, and the updated cell state in chronological order; Each memory block has three "gating" structures, including input gate, output gate and forget gate, which are used to protect and control the cell state, that is, to store and transmit time information; The forget gate is responsible for deciding how much of the cell state at the previous moment to retain in the cell state at the current moment, that is, deciding what information to discard from the cell state. The formula is: f t =σ(W f *[h t-1 ,x t ]+b f ) Where h t-1 and h t are the hidden layer states at time t-1 and t, respectively, x t-1 and x t are the input values ​​at time t-1 and t respectively, W f is the weight matrix, b f is the bias, σ is the sigmoid function; The input gate is responsible for deciding how much of the current input to the current cell state to retain. The formula is: i t =σ(W i *[h t-1 ,x t ]+b i ) Where i t is the input value of the input gate at time t, Create a new cell state value vector for the tanh layer at time t, W i and W C is the weight coefficient matrix, b i and b C is the bias term; Update the old cell state, the formula is: Where C t-1 and C t are the cell states at time t-1 and t respectively; The output gate is responsible for determining how many outputs the current cell state has. The formula is: the t =σ(W o *[h t-1 ,x t ]+b o ) h t =o t *fishy(C) t ) In the formula, o t is the output value of the output gate at time t, W o is the weight coefficient matrix, b o is the bias term.

4. The method for decomposing and analyzing air conditioning load for public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network according to claim 1 is characterized in that: The method of establishing a feature attention mechanism on the input side of the non-air conditioning load forecasting model, quantifying the correlation between input feature variables, dynamically adjusting the input feature weights, and enhancing the guidance of strongly correlated feature variables is as follows; On the input side, a feature attention mechanism is established to explore the correlation between the input features of weather, time, and historical load and the non-air conditioning load value at the corresponding moment. The attention weight of each input feature is dynamically allocated to focus on the impact of important feature variables and reduce the information of weakly related or redundant features. In order to obtain the correlation between each influencing feature and the current non-air-conditioning load, the feature attention mechanism is used to encode the quantized weights; After establishing the feature attention mechanism, the input of the CNN-LSTM network is no longer the original input feature value, but the weighted input feature value considering the correlation degree; Through the feature attention mechanism, the contribution rate of each input feature is adaptively extracted to improve the prediction accuracy.

5. The method for decomposing and analyzing the air conditioning load for public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network according to claim 4 is characterized in that: The method of encoding quantized weights using the feature attention mechanism is: The input features at the current moment and the hidden layer state h at the previous moment t-1 and memory information t-1 As the input of the feature attention mechanism, the attention weight of each input feature at the current moment is calculated respectively, and quantized using the calculation method of the multi-layer perceptron. The weight calculation is as follows: Where V e 、W e and U e is the multi-layer perceptron weight matrix for calculating attention weights, b e is the bias term; Use the Softmax function to Normalize the feature attention weights so that the sum is 1, as shown in the following formula: The characteristic correlation coefficient The corresponding input affects the eigenvalue Multiply to get the associated features that consider the contribution rate of different inputs to the features Hidden state h t Then update as follows: Among them, f CNN-LSTM is the LSTM network unit.

6. The method for decomposing and analyzing the air conditioning load for public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network according to claim 1 is characterized in that: The method of establishing a temporal attention mechanism on the output side of the non-air-conditioning load forecasting model, capturing the dependency of temporal information, dynamically adjusting the output of the hidden state, and enhancing the information expression of the key time step is as follows; A temporal attention mechanism is established on the decoding output side to obtain the hidden state weights of the corresponding time series, and to determine the degree of influence of the past time series state on the current non-air-conditioning load output, thereby strengthening the influence of strongly correlated time state information. In order to obtain the correlation between the historical time series information and the current output non-air conditioning load value, the temporal attention mechanism is used to decode the quantized weights; Establishing a dual attention mechanism Based on the original CNN-LSTM network, a feature attention mechanism is introduced before the first LSTM layer, and a temporal attention mechanism is introduced after the last LSTM layer. That is, the threshold load and load influencing factors in the non-air-conditioning season pass through the input layer, convolution layer, pooling layer, feature attention mechanism, LSTM layer 1, LSTM layer 2, LSTM layer 3, temporal attention mechanism, fully connected layer and output layer in sequence to obtain the non-air-conditioning load in the air-conditioning season; LSTM layer 1, LSTM layer 2 and LSTM layer 3 are LSTM layers, and the convolution layer, pooling layer and fully connected layer are CNN layers.

7. The method for decomposing and analyzing air conditioning load for public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network according to claim 5 is characterized in that: The method of decoding quantized weights by the temporal attention mechanism is: The hidden layer state h containing the time series information at each historical moment τ and the hidden layer state h at the previous moment of historical moment t t and memory information t As the input of the temporal attention mechanism, τ = 1, 2, …, t, …, T, the attention weights of historical moment t and each historical moment are calculated respectively, and quantized using the calculation method of the multi-layer perceptron. The weight is calculated as follows: In the formula, V d and W d is the multi-layer perceptron weight matrix for calculating attention weights, U d is the bias term; Use the Softmax function to Normalize the weights so that the sum of the temporal attention weights is 1, as shown in the following formula: Quantify the importance of the state information of the τth hidden layer to the predicted output at time t, and set all the weights The weighted summation of the corresponding hidden layer state information is used to obtain the comprehensive information c related to the temporal state characteristics. t , as shown below: The non-air-conditioning load value y to be predicted at time t t Comprehensive information related to temporal state characteristics c t After fusion through a single-layer feedforward network, the input of the CNN-LSTM network at each time t is obtained Where, and The weights and biases of a single-layer feedforward network are calculated to fusion input respectively; h t The CNN-LSTM network in the following formula iterates T times from the initial moment to the T moment: The hidden layer state h at the final moment T T and comprehensive information T The single-layer feedforward network is used for integration and finally the dimension transformation is performed by the single-layer feedforward network to obtain the predicted value of the non-air-conditioning load to be predicted. As shown in the following formula: Where W y and b w are the weights and biases of the single-layer feedforward network that integrates the output, V y and b y are the weights and biases of the feedforward network for dimension transformation, respectively.

8. The method for decomposing and analyzing air conditioning loads for public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network according to claim 5, characterized in that: Based on the seasonal operation characteristics of public building loads, the modeling method for establishing a public building air conditioning load decomposition model based on the dynamic difference method is as follows: The building is divided into air-conditioning season and non-air-conditioning season according to whether the air conditioner is turned on or not, and the following constraints are established: P 非空调季关口负荷 ≈P 非空调季非空调负荷 P 空调季关口负荷 =P 空调季非空调负荷 +P 空调季空调负荷 Establish a method to dynamically predict the non-air-conditioning load in the air-conditioning season by using the threshold load in the non-air-conditioning season; Due to the following constraints: P 空调负荷 =P 关口负荷 -P 非空调负荷 For the air-conditioning operation days divided by clustering, a public building air-conditioning load decomposition method is established, which uses the real-time measured gateway load actual value and the non-air-conditioning load prediction value dynamically predicted by the neural network to decompose the dynamic prediction value of the air-conditioning load. Based on the non-air-conditioning load prediction result of the air-conditioning season, the air-conditioning load prediction value of the air-conditioning season is obtained by real-time difference.

9. An analysis device for a decomposition and analysis method of air conditioning load for public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network, characterized in that: The analysis device includes: a clustering algorithm module, a module for establishing a model for distinguishing between air-conditioning operation days and non-operation days; The air conditioning non-operation day module is used as the training set for the neural network prediction model; The air conditioning operation day module serves as the model's test set. A method for deriving the air conditioning load of public buildings is developed by dynamically subtracting the air conditioning operation day module's threshold load curve from the non-air conditioning load curve obtained by the non-air conditioning load prediction model. A feature attention mechanism is established on the input side of the non-air conditioning load prediction model module. This module quantifies the correlation between input feature variables, dynamically adjusting the input feature weights and enhancing the guidance of strongly correlated feature variables. The non-air-conditioning load forecasting model module establishes a temporal attention mechanism module on the output side of this module to capture the dependencies of temporal information, dynamically adjust the output of hidden states, and enhance the information expression of key time steps; Based on the results of distinguishing between the air-conditioning operation day module and the air-conditioning non-operation day module, a public building air-conditioning load decomposition module based on the dynamic difference method is established; The method for establishing a model for distinguishing between air-conditioning operation days and non-operation days based on the clustering algorithm is as follows: Based on the clustering algorithm, a clustering method for the threshold load curves of public building users was established to distinguish between days when the user's air conditioner is turned on and days when it is not in operation. The user's daily threshold load curve was input into the clustering algorithm and clustered into two clusters. Select a clustering algorithm and use the actual sample that is most similar to other samples in each cluster as the cluster center to improve the clustering algorithm's sensitivity to noise and outliers and its tendency to fall into local optimality; Using a clustering algorithm, the output of two different curves after classification corresponds to the dates representing the air conditioner operating days and non-operating days. This method is more accurate than distinguishing between air conditioner on / off days based on experience such as seasonal temperature changes. The method of selecting the clustering algorithm is: First, for a given cluster number of 2, two representative objects are randomly selected as the initial cluster centers, the distance between each remaining object and the representative object is calculated and assigned to the nearest cluster, and the corresponding clustering results are generated; Then, the iterative process begins. For each iteration, a randomly selected non-center point replaces one of the original center points and the clustering results are recalculated; If the clustering effect is improved, the replacement is retained, otherwise the original center point is restored; when the replacement no longer improves the clustering effect, the iteration stops; The cost function is used to measure the quality of the clustering results. This function is used to measure the average dissimilarity between the object and the center point. The specific definition is as follows: Among them, p is a point in space, that is, a given object; o i Cluster C i The center point of the dataset; E is the sum of squared deviations of all objects in the dataset.

10. The analysis device for the decomposition and analysis method of air conditioning load for public buildings based on the two-stage attention mechanism fused with convolutional neural network and long short-term memory network according to claim 9 is characterized in that: The method of establishing the non-air conditioning load forecasting model module is: Establish the factors affecting the load of public buildings, including historical load, meteorological information and date type; The historical load includes the historical load of the past day, the load at the same time in the past week, and the load at the same time on the same characteristic day in the past month. Meteorological information includes temperature, humidity, wind speed and weather; Date types include time variables, weekday and rest day variables, and holiday variables; The load influencing factors of public building users are used as the input features of the training set and input into the CNN-LSTM network model through a sliding window. The non-air-conditioning load of public building users in the non-air-conditioning season is used as the target feature of the training set to train the CNN-LSTM network. Establish CNN-LSTM network model training samples; Among them, in the CNN convolutional neural network, the input layer, convolution layer, pooling layer, convolution layer, pooling layer, fully connected layer and output layer are connected in sequence; Using the trained model, the load influencing factors at the corresponding time of the air-conditioning season and the non-air-conditioning load of a typical day in the non-air-conditioning season are input from the input layer in the form of a sliding window. The trained model parameters are obtained through the CNN layer and the three-layer LSTM layer. The predicted non-air-conditioning load values ​​at each time of each day in the air-conditioning season are output through the output layer.

11. The analysis device for the decomposition and analysis method of air conditioning load for public buildings based on a two-stage attention mechanism fused with a convolutional neural network and a long short-term memory network according to claim 9, characterized in that: The public building air conditioning load decomposition module is a method for obtaining the public building air conditioning load by dynamically subtracting the air conditioning operation day threshold load curve from the non-air conditioning load prediction model.

12. An electronic device, characterized in that: The invention comprises one or more processors for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors execute any one of claims 1 to 8, which is a method for decomposing and analyzing the air-conditioning load for public buildings based on a two-stage attention mechanism that integrates a convolutional neural network and a long short-term memory network.

13. A storage medium storing a computer program, wherein: in, The computer program is configured to execute, when running, any one of claims 1-8, a method for decomposing and analyzing the air-conditioning load for public buildings based on a two-stage attention mechanism integrating a convolutional neural network and a long short-term memory network.

Images