Multi-step building cold load prediction method based on deep time-space feature fusion learning

Abstract

The invention provides a multi-step building cold load prediction method based on deep time-space feature fusion learning, and the method comprises the following steps: S1, obtaining time-space historical data related to building cold load prediction, including outdoor climate condition data, building cold load data, cold source system topological structure information and control state information; s2, data cleaning and data preprocessing are carried out on the existing historical time sequence operation data, the data cleaning comprises anomaly detection processing, and the data preprocessing comprises data smoothing and normalization processing; s3, establishing a multi-step building cold load prediction model CRG-Informer of depth time-space feature fusion learning based on existing data, wherein the model comprises a graph model establishment and embedding process and an Informer model prediction process; and S4, building cold load prediction is carried out by using the trained CRG-Informer model and real-time building operation and maintenance data. According to the method, the accuracy of building cold load prediction can be improved, and the building energy-saving effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and building cooling load prediction, and in particular to a multi-step building cooling load prediction method using deep temporal and spatial feature fusion learning. Background Art

[0002] With the continuous development of the global population, economy and urbanization, more and more buildings are being built and put into use, which makes energy shortage and environmental pollution problems increasingly prominent. The building operation and maintenance process consumes a lot of social energy and causes a lot of carbon emissions. In recent years, the organic combination of big data, artificial intelligence and the construction industry has shown huge energy-saving potential and can effectively reduce energy waste.

[0003] The key to successfully achieving stable, efficient and energy-saving operation of HVAC systems lies in how to obtain the demand load of the building and balance it with the supply load. However, complex and time-varying external meteorological conditions and internal object activities bring dynamics and uncertainty to the building load, which makes it difficult to analyze future trends and apply appropriate control strategies through engineering experience. In the prior art, the prediction of the cooling load of a building usually only takes into account the time-related building environment characteristics, such as temperature and humidity, which makes the prediction accuracy low, resulting in poor building energy-saving effects. Therefore, there is an urgent need for a method to further improve the accuracy of building cooling load prediction and thus improve building energy-saving effects. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a multi-step building cooling load prediction method based on deep temporal and spatial feature fusion learning, which can improve the accuracy of building cooling load prediction and enhance the energy-saving effect of the building.

[0005] The above technical objectives of the present invention are achieved through the following technical solutions:

[0006] A multi-step building cooling load prediction method based on deep temporal and spatial feature fusion learning, characterized by comprising the following steps:

[0007] S1, obtaining the temporal and spatial historical data related to the building cooling load prediction, including outdoor climate condition data, building cooling load data, cooling source system topology information and control status information;

[0008] S2, data cleaning and data preprocessing of the existing historical time series operation data, where data cleaning includes anomaly detection processing, and data preprocessing includes data smoothing and normalization processing;

[0009] S3, based on the existing data, establishes a multi-step building cooling load prediction model CRG-Informer with deep spatiotemporal feature fusion learning, including the graph model establishment and embedding process, and the Informer model prediction process;

[0010] S4 uses the trained CRG-Informer model and real-time building operation and maintenance data to predict the building cooling load.

[0011] Furthermore, the outdoor climate condition data in step S1 include ambient temperature and ambient humidity; the building cooling load data is calculated by multiplying the chilled water inlet and return water temperature difference with the refrigeration side flow rate and the refrigeration side liquid specific heat; the cold source system topology information includes the rated parameters and connection transfer relationship of the chiller, water pump and cooling tower; the control status information is the operating status of each device, including the chiller operating current ratio, the water pump operating frequency ratio and the cooling tower wind speed ratio.

[0012] Furthermore, the anomaly detection processing in step S2 includes missing value compensation and threshold anomaly detection; a forward filling method is used to compensate for missing values, and a sliding window threshold judgment is used to analyze abnormal data points in the threshold anomaly detection process.

[0013] Furthermore, the data smoothing in step S2 adopts a simple exponential smoothing method, and only the detected abnormal data points are processed.

[0014] Furthermore, the graph model building and embedding process in step S3 includes:

[0015] First, the spatial structure, control characteristics and equipment attributes of the cooling source system are used to establish the time series cooling source system spatial control relationship graph set G t =(V t ,E t ),in Represents a graph node, which represents four types of entity nodes: chillers, cooling towers, water pumps, and the building itself; E t Represents a directed edge, indicating the pipe connection and flow transfer relationship in the cold source system; the characteristics of the chiller node include the chiller operating current ratio and rated cooling capacity, the characteristics of the pump node include the pump operating frequency ratio and the pump rated flow, the characteristics of the cooling tower node include the operating wind speed ratio and the cooling tower rated flow, and the building node includes the cooling load and building area;

[0016] Then, the graph neural network is used to embed the graph model of the time series into a representation vector. The process can be expressed as:

[0017]

[0018] Among them, H (j+k)It is represented as a representation vector after j graph convolutions and k graph poolings. and are the initial representations of the i-th node and its edge nodes, respectively. Gconv(·) represents the graph convolution layer, σ represents the ReLU activation function, and Gpool(·) represents the graph pooling layer.

[0019] Finally, a three-layer perceptron model is used to get the final graph embedding vector: Z t =MLP(H (j+k) ).

[0020] Furthermore, the Informer model prediction process in step S3 includes:

[0021] Before inputting the Informer model, the graph embedding vector is fused with the outdoor climate condition data and building cooling load data in step S1, and the location embedding PE is added pos and global timestamp SE pos , the input vector of Informer can be expressed as:

[0022]

[0023] Among them, Conv1d(·) represents one-dimensional convolution, L x represents the length of the input vector, i∈{1,…,L x}, pos represents the input position information;

[0024] The probabilistic sparse self-attention mechanism, distillation mechanism and fully connected output mechanism are also introduced in the Informer model.

[0025] Furthermore, the probabilistic sparse self-attention mechanism:

[0026] Using the probability form to characterize the self-attention mechanism, the original self-attention mechanism can be expressed as:

[0027]

[0028] Using the probabilistic form, the self-attention of the i-th q can be expressed as:

[0029]

[0030] Where Q, K, and V represent the Query, Key, and Value matrices respectively, and q, k, and v represent the corresponding vectors respectively. Softmax is the activation function, and d represents the dimension of the key.

[0031]

[0032] Where p(·) represents probability, exp(·) represents exponential, and the importance of each query is measured by removing the KL divergence of the constant term. The importance measurement function can be expressed as

[0033]

[0034] The larger the importance value M calculated for each query, the more valuable it is. Finally, only the best u Qs are selected for dot product calculation.

[0035] Further, the distillation mechanism:

[0036] Not all qk pairs are used in the previous steps, and redundant items will be generated when the remaining qk pairs are multiplied by V. Therefore, a maximum pooling layer is added between adjacent probabilistic sparse self-attention mechanisms for downsampling, reducing the input sequence to 1 / 2 of the original to highlight the dominant features and reduce the complexity of the long sequence output. The distillation process from the jth layer to the j+1th layer can be expressed as:

[0037]

[0038] in[·] AB represents the attention block of Informer, and ELU(·) is the exponential linear activation function.

[0039] Furthermore, the fully connected output mechanism: the decoder structure of Informer is set to the standard structure, and the inference output value is synchronized through a fully connected layer.

[0040] The present invention has the following beneficial effects:

[0041] 1. This method not only takes into account the time-related building environment characteristics, but also proposes a modeling method for the control characteristic diagram of the cold source system that is related to both space and time, and uses graph neural networks to extract spatial features to improve the prediction effect from the perspective of feature processing.

[0042] 2. This method combines the Informer model suitable for time series forecasting tasks to extract long-term dependencies between graph embedding vectors and temporal information such as environmental loads, thereby improving the prediction effect from the perspective of model construction.

[0043] 3. This method requires less computing resources and can support multi-step, long-term cooling load forecasting under more hardware conditions, which is more conducive to the control optimization and overall coordination of the cooling source system. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flowchart of the specific steps of the present invention;

[0045] Figure 2The algorithm network diagram for deep spatiotemporal feature fusion learning and cooling load prediction of the present invention;

[0046] Figure 3 A schematic diagram of the topological structure of a building cooling source system in an embodiment of the present invention;

[0047] Figure 4 1 is a cross-sectional diagram of climate conditions and cooling load data collected in an embodiment of the present invention.

[0048] Figure 5 It is a mean square error line graph under different prediction conditions in the embodiment of the present invention.

[0049] Figure 6 This is a prediction error distribution diagram of the CRG-Informer algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0051] A multi-step building cooling load prediction method based on deep temporal and spatial feature fusion learning, such as Figure 1 As shown, the following steps are included:

[0052] S1: Obtain the spatiotemporal historical data related to the building cooling load prediction, including outdoor climate condition data, building cooling load data, cold source system topology information and control status information. Among them, the outdoor climate condition data includes ambient temperature and ambient humidity; the building cooling load data is calculated by multiplying the chilled water inlet and return water temperature difference with the flow rate on the freezing side and the specific heat of the liquid on the freezing side; the cold source system topology information mainly involves the rated parameters and connection transmission relationship of the three key equipments: chillers, water pumps and cooling towers; the control status information is the operating status of each device. Figure 4 Shown is a cross-sectional diagram of climate conditions and cooling load data collected in an embodiment of the present invention.

[0053] like Figure 3 The topological structure diagram of the building cooling source system of this embodiment is shown, which includes 3 chillers with rated cooling capacities of 1932, 1135, and 363.7 kW, respectively. The chiller operating current ratio is used to characterize its control state, which is a continuous value ranging from 0.4 to 1. There are 9 cooling water pumps with rated flow rates of 399.6, 399.6, 198, 198, 80, 450, 450, 259.2, and 259.2 m 3 / h, the pump operating frequency ratio (compared with the power frequency 50Hz) is used to characterize its control state, with a continuous value ranging from 0.7 to 1. There are 4 cooling towers with a rated air volume of 90km 3 / h, the cooling tower wind speed ratio is used to characterize its control state, and the range is discrete 0 or 1.

[0054] S2: Data cleaning and data preprocessing are performed on the existing historical time series operation data, wherein data cleaning includes anomaly detection processing, i.e., missing value compensation and threshold anomaly detection, and data preprocessing includes data smoothing and normalization processing. Anomaly detection processing mainly includes using the forward filling method to compensate for missing values, and using the sliding window threshold judgment to analyze abnormal data points in the threshold anomaly detection process. In this embodiment, the threshold is set to 20% to identify abnormal values, that is, a sudden rise or fall of data in any time window exceeding 20% ​​of the average value is judged to be abnormal; the data smoothing adopts a simple exponential smoothing method, and the process can be expressed as: in and Respectively represent the smoothed data at time t and time t-1, y t represents the data before smoothing at time t, and α represents the smoothing coefficient. In this example, α = 0.6. It is worth noting that the data smoothing in this method is not performed on the complete time series, but only on the detected abnormal data points to alleviate the information leakage problem caused by data smoothing; the normalization method uses maximum and minimum normalization to make the existing features between 0 and 1.

[0055] S3: Establish a multi-step building cooling load prediction model CRG-Informer based on deep temporal and spatial feature fusion learning based on existing data. Figure 2 The figure shows the algorithm network diagram of the deep spatiotemporal feature fusion learning and cooling load prediction of this method. The process can be divided into S3-1 graph model establishment and embedding and S3-2 Informer model prediction.

[0056] S3-1 Graph model establishment and embedding process:

[0057] First, the spatial structure, control characteristics and equipment attributes of the cooling source system are used to establish the time series cooling source system spatial control relationship graph set G t =(V t ,E t ),in Represents a graph node, which represents four types of entity nodes: chillers, cooling towers, water pumps, and the building itself. t Represents a directed edge, indicating the pipe connection and flow transfer relationship in the cold source system. The characteristics of each node include the control characteristics and device attributes of the corresponding equipment. The characteristics of the chiller node include the chiller operating current ratio and rated cooling capacity, the characteristics of the pump node include the pump operating frequency ratio and rated flow, the characteristics of the cooling tower node include the operating wind speed ratio and the cooling tower rated flow, and the building node includes the cooling load and building area.

[0058] Then, the graph neural network is used to embed the graph model of the time series into a representation vector. The process can be expressed as:

[0059]

[0060] Among them, H(j+k) represents the representation vector after j graph convolutions and k graph pooling. and are the initial representations of the i-th node and its edge nodes, respectively. Gconv(·) represents the graph convolution layer, σ represents the ReLU activation function, and Gpool(·) represents the graph pooling layer.

[0061] Finally, a three-layer perceptron model is used to get the final graph embedding vector: Z t =MLP(H (j+k) ).

[0062] S3-2 Informer model prediction process:

[0063] Before inputting into the Informer model, the graph embedding vector is combined with the outdoor climate condition data and building cooling load data in step S1. Fusion and adding location embedding PE pos and global timestamp SE pos , the input vector of Informer can be expressed as:

[0064]

[0065] Among them, Conv1d(·) represents one-dimensional convolution, L x represents the length of the input vector, i∈{1,…,L x}, pos represents the input position information; t and t-1 refer to time t and time t-1, and p represents that there are p types of global timestamps.

[0066] like Figure 2 As shown in the figure, Informer contains Ne encoder modules and Nd decoder modules. Compared with the traditional Transformer model, the Informer model of this method introduces a probabilistic sparse self-attention mechanism, a distillation mechanism, and a fully connected layer to better adapt to time series prediction tasks.

[0067] Introducing the probabilistic sparse self-attention mechanism: It uses the probabilistic form to characterize the self-attention mechanism. The original self-attention mechanism can be expressed as:

[0068]

[0069] Using the probabilistic form, the self-attention of the i-th q can be expressed as:

[0070]

[0071] Where Q, K, and V represent the Query, Key, and Value matrices respectively, and q, k, and v represent the corresponding vectors respectively. Softmax is the activation function, and d represents the dimension of the key.

[0072]

[0073] Where p(·) represents probability, exp(·) represents exponential, and the importance of each query is measured by removing the KL divergence of the constant term. The importance measurement function can be expressed as:

[0074]

[0075] The larger the importance value M calculated for each query, the more valuable it is. Finally, only the best u Qs are selected for dot product calculation.

[0076] Introducing distillation mechanism: Since not all qk pairs are used in the previous steps, a large number of redundant items will be generated when the remaining qk pairs are multiplied by V. Therefore, a maximum pooling layer is added between adjacent probabilistic sparse self-attention mechanisms for downsampling, reducing the input sequence to 1 / 2 of the original to highlight the dominant features and reduce the complexity of long sequence output. The distillation process from the jth layer to the j+1th layer can be expressed as:

[0077]

[0078] in[·] AB represents the attention block of Informer, and ELU(·) is the exponential linear activation function.

[0079] In addition, the robustness of the encoder is enhanced by constructing several shallower stack copies with half the input, and the number of stack copies can be set arbitrarily. In this embodiment, two additional stack copies are constructed, and finally the connection of all stacks is output as the depth representation of the encoder.

[0080] Introducing a fully connected layer: Set the decoder structure of Informer to a standard structure, and finally synchronize the inference output value through a fully connected layer. Compared with the dynamic decoding in Transformer, it reduces the error caused by step-by-step time prediction inference and reduces the inference time.

[0081] Figure 2The Informer prediction model in can be summarized as follows: (1) As shown in the lower left corner of the figure, the input vector of position embedding and feature fusion is added, and then enters the encoder module on the left. It is normalized in the middle and divided into Q, K, V. The probability sparse self-attention calculation is performed and added to the value before normalization. After distillation, normalization, and full connection layer. At this time, after one encoding process, there are a total of Ne. And there are two stack copies of the input halved, as shown by the red arrow and blue arrow in the figure. (2) As shown in the figure, after all the encoding processes on the left are spliced ​​together, they are put into the multi-head attention on the right. The predicted value is in the lower right corner. The value before 0 is known in the past, and the value after 0 is to be predicted. Then it enters the decoder module and is finally output through the fully connected layer.

[0082] S4, use the trained CRG-Informer model and real-time building operation and maintenance data to predict the building cooling load. In this embodiment, the data set is divided in chronological order, with the first 70% as the training set and the last 30% as the test set. In order to verify the superiority and applicability of this method, this embodiment also compares the prediction results of four different data collection intervals: 5 minutes, 10 minutes, 15 minutes, 30 minutes, and 5 different prediction steps: 6 steps, 12 steps, 24 steps, 72 steps, and 144 steps. At the same time, five major time series prediction models including LSTM, LSTNet, GRU, XGBoost, and Informer are selected as benchmark comparisons, and the mean absolute error MAE and mean square error MSE are selected as evaluation indicators of the model.

[0083] Among them, LSTM: Long Short-Term Memory, long short-term memory neural network; GRU: gated recurrent unit, gated recurrent unit (neural network); LSTNet: Long-and Short-term Time-series Networks, long-term and short-term time series network; XGBoost: eXtreme Gradient Boosting, extreme gradient boosting.

[0084] The prediction results of the above methods under different prediction conditions can be seen in Table 1. The optimal value in each case is indicated in bold. The last row indicates the optimal number of each algorithm. Tables 2 and 3 show the training time of each algorithm under different data intervals (data volume) in 24-step prediction and 144-step prediction, respectively. Figure 5The mean square error of the above method under different prediction conditions is shown. It can be found that the CRG-Informer algorithm proposed in this method performs well in various situations, and as the prediction step increases, the error increases evenly and slowly. In the 5-minute prediction interval, compared with the optimal value in the baseline model, the mean square error is reduced by 32.0% (6-step prediction), 14.9% (12-step prediction), 7.88% (24-step prediction), 22.8% (72-step prediction), and 23.5% (144-step prediction). At the same time, the proposed method has obvious resource advantages in long-term prediction. As the prediction step increases, the time required for model training only increases slightly, which is much better than typical models based on recurrent neural networks (RNN) and convolutional neural networks (CNN). Appendix Figure 6 It shows the error distribution of CRG-Informer in various prediction situations. Overall, its positive and negative deviations are relatively uniform, especially in short-term and short-step predictions, which also proves the good prediction performance of the model.

[0085] Table 1 Cooling load prediction results under various prediction conditions

[0086]

[0087] Table 2 Model training time under 24-step cooling load prediction

[0088]

[0089] Table 31 Model training time under 44-step cooling load prediction

[0090]

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A multi-step building cooling load prediction method based on deep temporal and spatial feature fusion learning, characterized by: The following steps are involved: S1, obtaining the temporal and spatial historical data related to the building cooling load prediction, including outdoor climate condition data, building cooling load data, cooling source system topology information and control status information; S2, data cleaning and data preprocessing of the existing historical time series operation data, where data cleaning includes anomaly detection processing, and data preprocessing includes data smoothing and normalization processing; S3, based on the existing data, establishes a multi-step building cooling load prediction model CRG-Informer with deep spatiotemporal feature fusion learning, including the graph model establishment and embedding process, and the Informer model prediction process; S4 uses the trained CRG-Informer model and real-time building operation and maintenance data to predict the building cooling load.

2. According to the multi-step building cooling load prediction method of deep temporal and spatial feature fusion learning according to claim 1, it is characterized by: The outdoor climate condition data in step S1 include ambient temperature and ambient humidity; the building cooling load data is calculated by multiplying the chilled water inlet and return water temperature difference with the freezing side flow rate and the freezing side liquid specific heat; the cold source system topology information includes the rated parameters and connection transfer relationship of the chiller, water pump and cooling tower; the control status information is the operating status of each device, including the chiller operating current ratio, water pump operating frequency ratio and cooling tower wind speed ratio.

3. According to the multi-step building cooling load prediction method of deep temporal and spatial feature fusion learning in claim 1, it is characterized by: The anomaly detection process in step S2 includes missing value compensation and threshold anomaly detection; a forward filling method is used to compensate for missing values, and a sliding window threshold judgment is used to analyze abnormal data points in the threshold anomaly detection process.

4. The multi-step building cooling load prediction method based on deep temporal and spatial feature fusion learning according to claim 3 is characterized by: The data smoothing in step S2 adopts a simple exponential smoothing method, and only the detected abnormal data points are processed.

5. The multi-step building cooling load prediction method based on deep temporal and spatial feature fusion learning according to claim 1 is characterized by: The graph model building and embedding process in step S3 includes: First, the spatial structure, control characteristics and equipment attributes of the cooling source system are used to establish the time series cooling source system spatial control relationship graph set G t =(V t ,E t ),in Represents a graph node, which represents four types of entity nodes: chillers, cooling towers, water pumps, and the building itself; E t Represents a directed edge, indicating the pipe connection and flow transfer relationship in the cold source system; the characteristics of the chiller node include the chiller operating current ratio and rated cooling capacity, the characteristics of the pump node include the pump operating frequency ratio and the pump rated flow, the characteristics of the cooling tower node include the operating wind speed ratio and the cooling tower rated flow, and the building node includes the cooling load and building area; Then, the graph neural network is used to embed the graph model of the time series into a representation vector. The process can be expressed as: Among them, H (j+k) It is represented as a representation vector after j graph convolutions and k graph poolings. and are the initial representations of the i-th node and its edge nodes, respectively. Gconv(·) represents the graph convolution layer, σ represents the ReLU activation function, and Gpool(·) represents the graph pooling layer. Finally, a three-layer perceptron model is used to get the final graph embedding vector: Z t =MLP(H (j+k) ).

6. The multi-step building cooling load prediction method based on deep temporal and spatial feature fusion learning according to claim 5 is characterized by: The Informer model prediction process in step S3 includes: Before inputting the Informer model, the graph embedding vector is fused with the outdoor climate condition data and building cooling load data in step S1, and the location embedding PE is added pos and global timestamp SE pos , the input vector of Informer can be expressed as: Among them, Conv1d(·) represents one-dimensional convolution, L x represents the length of the input vector, i∈{1,…,L x }, pos represents the input position information; The probabilistic sparse self-attention mechanism, distillation mechanism and fully connected output mechanism are also introduced in the Informer model.

7. The multi-step building cooling load prediction method based on deep temporal and spatial feature fusion learning according to claim 6 is characterized by: Probabilistic Sparse Self-Attention Mechanism: Using the probability form to characterize the self-attention mechanism, the original self-attention mechanism can be expressed as: Using the probabilistic form, the self-attention of the i-th q can be expressed as: Where Q, K, V represent the Query, Key, and Value matrices respectively, and q, k, and v represent the corresponding vectors respectively, softmax is the activation function, and d represents the dimension of the key; and, Where p(·) represents probability, exp(·) represents exponential, and the importance of each query is measured by removing the KL divergence of the constant term. The importance measurement function can be expressed as: The larger the importance value M calculated for each query, the more valuable it is. Finally, only the best u Qs are selected for dot product calculation.

8. The multi-step building cooling load prediction method based on deep temporal and spatial feature fusion learning according to claim 7 is characterized by: Distillation mechanism: Not all qk pairs are used in the previous steps, and redundant items will be generated when the remaining qk pairs are multiplied by V. Therefore, a maximum pooling layer is added between adjacent probabilistic sparse self-attention mechanisms for downsampling, reducing the input sequence to 1 / 2 of the original to highlight the dominant features and reduce the complexity of the long sequence output. The distillation process from the jth layer to the j+1th layer can be expressed as: in[·] AB represents the attention block of Informer, and ELU(·) is the exponential linear activation function.

9. The multi-step building cooling load prediction method based on deep temporal and spatial feature fusion learning according to claim 8 is characterized by: Fully connected output mechanism: Set the decoder structure of Informer to the standard structure and synchronize the inference output value through a fully connected layer.

Images