A method, device, medium and equipment for predicting load of building air conditioning system

Abstract

The present invention discloses a method, device, medium and equipment for predicting the load of a building air conditioning system, and relates to the field of computer technology. First, the historical time series data of the cooling load of air conditioners at different positions in the building air conditioning system and the load impact data of the corresponding historical period are obtained as sample data, and then a bidirectional spatial convolutional network is improved based on the bidirectional temporal convolutional network. A cross-attention module is connected in series after the parallel structure formed by the bidirectional spatial convolutional network and the bidirectional long short-term memory network to construct a prediction model, and then the prediction model is trained by the sample data, and the hyperparameters are optimized by the crown porcupine optimization algorithm to obtain the load prediction model of the building air conditioning system, and finally the cooling load of each air conditioner in the building air conditioning system at the target time is predicted based on the data to be predicted by the load prediction model of the building air conditioning system. The present invention improves the accuracy of the prediction of the cooling load of air conditioners at different positions in the building air conditioning system.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method, device, medium and equipment for predicting load of a building air-conditioning system. Background Art

[0002] At present, building energy conservation and environmental protection are the key issues of today's society. In the field of construction, the use of efficient and energy-saving cooling systems is one of the key measures to improve building energy efficiency. Radiant cooling, as an efficient and comfortable cooling method, is being used more and more widely in buildings. However, radiant cooling systems have problems such as long response time, large inertia, sensitivity to environmental conditions, high condensation risk, and complex control strategies. To solve these problems, it is necessary to predict load changes through load forecasting, so as to optimize system operation, improve efficiency and comfort, and reduce operating costs and risks. Therefore, accurate load forecasting is the key to solving the problems of radiant cooling air-conditioning systems and achieving efficient and energy-saving operation.

[0003] In the prior art, air conditioning load forecasting methods usually adopt data-driven technology, which uses historical data of complex nonlinear systems to predict air conditioning load. This method aims to extract effective features and patterns from past operating data to achieve accurate load forecasting.

[0004] However, existing data-driven air-conditioning load forecasting methods lack the ability to model spatial characteristics. Most methods only process the load time series of a single air conditioner (such as the air-conditioning load data of a room), while ignoring the spatial relationship between different locations in the building, resulting in poor load forecasting accuracy for the air-conditioning system in the building. Summary of the invention

[0005] Based on this, it is necessary to provide a method, device, medium and equipment for load prediction of a building air conditioning system in response to the above technical problems.

[0006] The present invention adopts the following technical solutions:

[0007] The present invention provides a load prediction method for a building air-conditioning system. The method first obtains historical time series data of cooling loads of air conditioners at different positions in the building air-conditioning system and load influence data of corresponding historical periods as sample data, then improves a bidirectional spatial convolutional network based on a bidirectional temporal convolutional network to obtain a bidirectional spatial convolutional network, connects a cross-attention module in series after a parallel structure formed by the bidirectional spatial convolutional network and a bidirectional long short-term memory network, and constructs a prediction model. Then, the prediction model is trained through the sample data, and hyperparameters are optimized through a crown porcupine optimization algorithm to obtain a load prediction model for the building air-conditioning system. Finally, the load prediction model for the building air-conditioning system extracts spatial distribution characteristics of time series data of cooling loads of air conditioners at different positions and influence relationship characteristics between air conditioners and load influence data based on the data to be predicted, and extracts the time series change characteristics of the data to be predicted, so as to fuse the three through a cross-attention module to obtain a fusion feature, so as to predict the cooling load of each air conditioner in the building air-conditioning system at a target time according to the fusion feature.

[0008] The present invention provides a load prediction device for a building air conditioning system, comprising:

[0009] An acquisition module is used to acquire historical time series data of cooling loads of air conditioners at different locations in the building air conditioning system and load impact data of corresponding historical periods as sample data;

[0010] A construction module is used to convert the extended causal convolution layer in the bidirectional temporal convolutional network into a 2D convolution to obtain a bidirectional spatial convolutional network, and connect a cross-attention module to the parallel structure formed by the bidirectional spatial convolutional network and the bidirectional long short-term memory network to build a prediction model;

[0011] A training module is used to train the prediction model using the crown porcupine optimization algorithm according to sample data to obtain a load prediction model for the building air conditioning system;

[0012] The prediction module is used to input the data to be predicted into the load prediction model of the building air-conditioning system, extract the spatial distribution characteristics of the time series data of the air-conditioning cooling load at different locations and the influence relationship characteristics of various load influence data and the time series data of the air-conditioning cooling load at different locations through a bidirectional spatial convolutional network, extract the time series change characteristics of the time series data of the air-conditioning cooling load at different locations and various load influence data in the data to be predicted through a long short-term memory neural network, and fuse the time series change characteristics, spatial distribution characteristics and influence relationship characteristics through a cross-attention module to obtain a fusion feature, so as to determine the cooling load of each air conditioner in the building air-conditioning system at the target time according to the fusion feature.

[0013] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned building air conditioning system load prediction method is implemented.

[0014] The present invention provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned building air conditioning system load prediction method when executing the program.

[0015] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0016] The flow of cold air and temperature changes between different areas in a building are usually closely related to the load data of air conditioners at various locations in the air-conditioning system of the building. The present invention converts the extended causal convolution layer in the bidirectional temporal convolutional network into a 2D convolution to obtain a bidirectional spatial convolutional network. The present invention can effectively capture the spatial distribution characteristics of air-conditioning load data at different locations in the air-conditioning system, as well as the influence relationship characteristics of various load influence data and the time series data of air-conditioning cold load at different locations through the bidirectional spatial convolutional network. Therefore, the coupling relationship of load data changes between air conditioners at different locations in the building, as well as the coupling relationship between air-conditioning loads at different locations and load influence data are reflected through spatial distribution characteristics and influence relationship characteristics, so that the prediction model can more comprehensively learn the spatial dependence of each air-conditioning load change, and extract the time series change characteristics of air-conditioning cold load time series data at different locations and various load influence data. The features of the spatial dimension and the time dimension are integrated through the cross-attention module, thereby improving the grasp of the load data and load influence data of the air-conditioning system in the building in the spatiotemporal dimension, thereby improving the load prediction accuracy of the air-conditioning system in the building. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0018] Figure 1 A schematic diagram of a flow chart of a method for predicting load of a building air conditioning system provided by the present invention;

[0019] Figure 2 A schematic diagram of a BiSCN network structure provided by the present invention;

[0020] Figure 3 A schematic diagram of the storage unit structure of an LSTM provided by the present invention;

[0021] Figure 4 A schematic diagram of a BiLSTM network architecture provided by the present invention;

[0022] Figure 5 A schematic diagram of a BiSCN-BiLSTM-Attention topology structure provided by the present invention;

[0023] Figure 6 A schematic diagram of the process of a CPO-BiSCN-BiLSTM-Attention model provided by the present invention;

[0024] Figure 7 A schematic diagram of a building air conditioning system load prediction device provided by the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0026] At present, the existing air conditioning load forecasting methods lack the ability to model spatial features. For example, some studies use the Autoregressive Integrated Moving Average Model (ARIMA) model to model the load of a single room. However, ARIMA only relies on the historical data of the room and cannot consider factors such as the temperature and humidity of adjacent rooms, resulting in reduced prediction accuracy. Support Vector Regression (SVR) is applied to load forecasting in specific areas. SVR mainly relies on input features and has no spatial feature modeling, which makes it impossible to capture the transfer and coupling of cold air between multiple areas.

[0027] In addition, traditional long short-term memory networks (LSTM) or gated recurrent units (GRU) cannot effectively capture long-term dependencies, especially high-frequency and non-stationary load data. Many models assume that the inputs of all time steps contribute equally to the load and fail to dynamically adjust the model's attention to different time steps. For example, the machine learning library (eXtreme Gradient Boosting, XGBoost) is applied to load forecasting in certain areas. XGBoost relies on a fixed feature combination and cannot dynamically adjust the importance of time features. As the complexity of data increases (such as multi-regional loads and multiple meteorological parameters), simple models find it difficult to handle feature selection and learning of high-dimensional inputs. Many deep learning models are "black box" models, and it is difficult to explain the complex relationship between their prediction results and data. For example, when using deep neural networks to predict air conditioning loads, although deep neural networks may be better than traditional models in terms of prediction accuracy, their internal mechanisms are complex and users find it difficult to understand how the model derives its prediction results.

[0028] The improved Bidirectional Spatial Convolutional Network (BiSCN) of the present invention uses the convolution kernel range from the neighboring position to the global area to extract spatial correlation, which allows the model to not only focus on the load changes at one location, but also capture the spatial coupling characteristics across regions (such as the impact of cold air flow transmission). The bidirectional long short-term memory network (Bidirectional Long Short-Term Memory) allows the model to learn information from past time steps and capture the potential impact of future time steps, which solves the problem that the traditional unidirectional LSTM cannot fully model long-term and nonlinear trends. Through the joint use of the BiSCN module and the BiLSTM, the model can effectively extract features from different locations and multiple time steps, and automatically select features related to the prediction task, which improves the learning ability of complex multi-regional and multi-time period data. Through the output and weight visualization of the temporal attention mechanism, users can clearly see which time steps and regions have the greatest impact on load forecasting. This improves the interpretability of the model and enhances the user's trust in the prediction results.

[0029] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0030] Figure 1 The present invention is a flow chart of a method for predicting load of a building air conditioning system, which specifically includes the following steps:

[0031] S101: Acquire historical time series data of cooling loads of air conditioners at different locations in a building air conditioning system and load impact data of corresponding historical periods as sample data.

[0032] S102: Convert the extended causal convolution layer in the bidirectional temporal convolutional network into a 2D convolution to obtain a bidirectional spatial convolutional network, connect a cross-attention module in series after the parallel structure formed by the bidirectional spatial convolutional network and the bidirectional long short-term memory network, and build a prediction model.

[0033] S103: A prediction model is trained using the crested porcupine optimization algorithm according to the sample data to obtain a load prediction model for the building air conditioning system.

[0034] S104: Input the data to be predicted into the load prediction model of the building air-conditioning system, extract the spatial distribution characteristics of the time series data of the cooling load of the air conditioners at different locations and the influence relationship characteristics of various load influence data and the time series data of the cooling load of the air conditioners at different locations through a bidirectional spatial convolutional network, extract the time series change characteristics of the cooling load of the air conditioners at different locations and various load influence data in the data to be predicted through a long short-term memory neural network, fuse the time series change characteristics, spatial distribution characteristics and influence relationship characteristics through a cross-attention module to obtain a fusion feature, so as to determine the cooling load of each air conditioner in the building air-conditioning system at the target time according to the fusion feature.

[0035] For the convenience of explanation, the following description will only take the server as the execution subject. The server mentioned in the present invention can be a server set up on the business platform, or a device such as a desktop computer, a laptop computer, etc. that can execute the solution of the present invention.

[0036] Generally, when the server of the business platform needs to perform multi-dimensional load forecasting for the integrated energy system, it needs to first train the model with historical data, and then perform multi-dimensional load forecasting with the trained model.

[0037] In one or more embodiments of the present invention, the server needs to obtain the historical time series data of the cooling load of the air conditioners at different positions in the building air conditioning system, which can be the cooling load data of the air conditioners at different positions in the building air conditioning system in a certain historical period. Of course, if only the historical cooling load time series data is used for prediction, the variables are single and the high-dimensional characteristics of the data cannot be fully mined. The load impact data of the corresponding historical period can also be obtained as sample data. Specifically, in one or more embodiments of the present invention, the load impact data includes: outdoor temperature, indoor temperature, outdoor humidity, indoor humidity, solar radiation intensity, wind speed, holiday attribute information, real-time number of people in each area of ​​the building, and at least one of the water supply temperature and return water temperature of the air conditioning system. Among them, the holiday attributes can be encoded, for example, the working day is encoded as the number 1, and the holiday is encoded as the number 0. For various types of indoor load impact data, it can usually be the corresponding indoor data of each area of ​​the building. The area division in the building can be based on the location of different air conditioners in the air conditioning system.

[0038] Since there are many factors that affect the summer radiant cooling air-conditioning load, there are different coupling relationships between various load impact data. Selecting different load impact data will also have a great impact on the prediction accuracy. The selection of input characteristic parameters has an important influence on the accuracy of the prediction model. It is necessary to screen out a set of characteristic parameters with as little redundancy as possible but sufficient to explain the behavior of the output variables.

[0039] Based on this, in one or more embodiments of the present invention, in order to reduce the complexity of the model, cosine similarity can be used to determine the correlation between the building air-conditioning system and various load impact data, and then the various load impact data can be preliminarily screened. For example, if the cosine similarity is less than 0.4, it can be considered that the correlation is low and the corresponding data can be eliminated.

[0040] Specifically, the server can obtain the historical time series data of the cooling load of air conditioners in different locations in the building's air-conditioning system, and obtain multiple types of load impact data for the corresponding historical period, including outdoor temperature, indoor temperature, outdoor humidity, indoor humidity, solar radiation intensity, wind speed, holiday attribute information, real-time number of people in each area of ​​the building, and supply water temperature and return water temperature of the air-conditioning system.

[0041] Then, for the air conditioner at each location, the cosine similarity between the cooling load history time series data of the air conditioner at that location and the various load impact data is determined, so as to screen the various load impact data according to the preset similarity threshold, and determine one or more types of load impact data for load forecasting at that location. The formula for cosine similarity is as follows: .

[0042] In the formula, is the cosine similarity, For the building i The first position of the air conditioner k characteristic variables (each type of load impact data is a characteristic, such as indoor temperature, etc.), For the building i The cooling load historical time series data of each air conditioner.

[0043] In one or more embodiments of the present invention, the sample data includes the historical time series data of cooling load and load impact data of air conditioners at different locations. When performing the load impact data screening work, each air conditioner corresponding position forms a multidimensional time series tensor separately. ,in N 1 is the time step, d is the number of features at that location. The cosine similarity is calculated for all load impact data at each location and the historical time series data of cooling load, and the features with high similarity are screened out for the input of subsequent models.

[0044] After the screening, each location has its own characteristic variables, such as air-conditioning location 1: outdoor temperature, indoor temperature and outdoor humidity; air-conditioning location 2: outdoor temperature, indoor temperature, outdoor humidity, indoor humidity, solar radiation intensity and real-time number of people in the building area; the number of features after screening varies for different locations.

[0045] Furthermore, data preprocessing is one of the key steps to improve data quality and enhance model training efficiency, which usually includes data normalization and detection and correction of outliers.

[0046] In one or more embodiments of the present invention, the server may first normalize the sample data. The normalization of the sample data is to eliminate the influence of inconsistent dimensions between different parameters. The present invention uses Z-score standardization for the calculation method of normalization processing of each type of sample data, as shown in the following formula: .

[0047] in, X standard is the standardized data set, μ For the current i The first position of the air conditioner k The mean of the characteristic variables, σ For the current i The first position of the air conditioner k The standard deviation of the characteristic variable.

[0048] Then, outlier detection and correction work is performed. Outliers usually refer to one or more data values ​​in the data that are significantly different from other data values. In the prediction model of radiant cooling building air conditioning system, the existence of outliers makes the data set highly heteroscedastic, thereby reducing the prediction accuracy of the prediction model. Therefore, before model training, it is necessary to eliminate the outliers in the sample data. The present invention uses the Generalized Extreme Studentized Deviate (GESD) algorithm, which is currently widely used in building abnormal data detection, to detect outliers in the data set and remove data values ​​that are significantly different from other data values.

[0049] Afterwards, the preprocessed sample data can be input into the pre-built prediction model. In one or more embodiments of the present invention, the prediction model includes a bidirectional spatial convolutional network, a bidirectional long short-term memory network (Bidirectional LongShort-Term Memory, BiLSTM) and an attention mechanism (Attention), forming a BiSCN-BiLSTM-Attention neural network model, which is optimized by the Crested Porcupine Optimizer (CPO), and therefore can also be called a CPO-BiSCN-BiLSTM-Attention model.

[0050] In one or more embodiments of the present invention, the prediction model is used to capture the temporal variation characteristics and spatial distribution characteristics in the sample data, wherein the attention mechanism adopts the cross-attention mechanism, which can effectively integrate the temporal variation characteristics and spatial distribution characteristics in the historical time series data of the cooling load of the air conditioner and the load impact data of the corresponding historical period, and can calculate the attention between multiple input sequences or modes, so as to better capture and utilize the correlation in multimodal data. The crown porcupine optimization algorithm is used to optimize the structure and parameters of the proposed BiSCN-BiLSTM-Attention neural network model, so as to accurately predict the cooling load of the building air conditioning system, optimize the building energy consumption management, and help the building management system accurately control the operation time and intensity of the refrigeration system, and reduce unnecessary energy consumption.

[0051] Specifically, for the bidirectional spatial convolutional network, the bidirectional spatial convolutional network is based on the bidirectional temporal convolutional network (Bidirectional Temporal Convolutional Network, BiTCN) with a bidirectional temporal convolution structure. The extended causal convolution layer is one of the important components in BiTCN, which is used to extract the long-term dependencies of time series data. The bidirectional parallel structure is an important design in BiTCN, which is mainly used to capture forward (past to present) and reverse (future to present) features from time series data. This structure improves the prediction performance of the model, especially in tasks that require a comprehensive understanding of temporal dependencies.

[0052] By converting the extended causal convolution layer in BiTCN into a 2D convolution, it is improved to BiSCN. The 2D convolution operation applies the convolution kernel sliding to the input two-dimensional feature map, calculates the weighted sum of the local area, and forms the output feature map. This sliding window operation enables the model to extract patterns (such as the relationship between adjacent positions) in the local area and capture more complex global spatial distribution features through layer-by-layer stacking. By stacking multiple convolutional layers, the 2D convolution gradually expands the receptive field and captures spatial distribution features in a larger range.

[0053] The modified BiSCN combines the advantages of BiTCN's bidirectional processing and enhances the model's ability to learn spatial sequences in the data. In BiSCN, the idea of ​​future forward and past reverse is transformed into bidirectional data processing in the spatial dimension. Convolution operations are performed from left to right forward and from right to left reverse in the spatial dimension. Forward convolution uses 2D convolution to capture the local spatial distribution characteristics of air-conditioning loads at different locations in the load data set. Reverse convolution uses transposed convolution to restore low-resolution feature maps to higher resolutions while maintaining spatial structure and details. It can restore and enhance global details. This design can combine the local spatial information provided by forward convolution with the global details restored by reverse convolution, which helps to generate higher quality output. Figure 2 As shown, Figure 2 This is a schematic diagram of a BiSCN network structure in the present invention.

[0054] BiSCN adapts the input data to 2D convolution through the data conversion layer, captures the spatial distribution characteristics in the sequence through 2D convolution, restores and enhances global details through transposed convolution, stabilizes training through weight normalization and batch normalization, introduces nonlinearity through activation function (ReLU), prevents overfitting through dropout, uses 1x1 convolution in the residual connection part, mainly used to adjust the number of channels so that the number of channels of input and output are consistent, solves the gradient disappearance problem through residual connection, and enables the network to learn complex spatial sequence characteristics while maintaining a stable and efficient training process. Therefore, the BiSCN model designed by the present invention is suitable for tasks that require a detailed understanding of spatial relationships.

[0055] Usually, the air conditioning cooling load at a certain location is not only affected by the environmental conditions at that location, but also by the air conditioning cooling load and environmental conditions (such as the temperature of adjacent rooms) at other locations, and the mutual relationship between them is not necessarily unidirectional. Through the bidirectional convolution structure, the model can capture the interaction between different locations, not just the propagation relationship in one direction. Therefore, the bidirectional parallel structure of BiTCN is used to extract the spatial distribution characteristics of the historical time series data of the cooling load and the load impact data of the corresponding historical period, not from the front and back transmission of the time dimension, but to simulate the mutual influence between different locations in the spatial dimension. Specifically, it can be understood as aggregating spatial distribution characteristics in one direction and in reverse order.

[0056] Furthermore, in one or more embodiments of the present invention, after the types of load impact data corresponding to air conditioners at different locations are screened, the number and length of features extracted and screened at different locations are different. In order to better perform 2D convolution operations, a mask operation can be added here.

[0057] Specifically, the 2D convolution in the BiSCN model can be a 2D masked convolution. The server can mask the data of various load impact data other than the load impact data used for load prediction at the location for the air conditioner at each location through 2D masked convolution, thereby extracting the influence relationship characteristics of the masked various load impact data and the historical cooling load time series data of the air conditioners at different locations.

[0058] The mask is a matrix of the same shape as the convolution kernel, whose elements are usually 0 or 1. A 1 in the mask indicates that the data at that position should be considered, and a 0 indicates that the data at that position is ignored. When performing convolution calculations, the mask is multiplied element-by-element with the input feature variables to determine which input feature variables will be included in the calculation.

[0059] The input data of BiSCN can be a 4D tensor in the following format: , where B is the batch size, i.e., multiple time segments are regarded as a batch, C represents the historical time series data of cooling load at each location and various load influencing factors, H represents the time step, i.e., the length of the time series (such as 24 hours in a day), and W represents the number of locations or air conditioners (i.e., the number of locations corresponding to the number of air conditioners in the air conditioning system).

[0060] For the bidirectional long short-term memory network, first of all, LSTM (Long Short-Term Memory) is a special recurrent neural network (RNN) unit, which solves the gradient vanishing and gradient exploding problems of traditional RNN by introducing forget gate, input gate and output gate. The input gate determines which information in the input sequence can be passed into the unit, the forget gate selectively forgets part of the information of the hidden state of the previous unit, and the output gate filters the information passed into the next unit. The basic mathematical formulas of these gates are as follows: , , , , .

[0061] in, X LSTMt express t Input data of time steps. For BiLSTM network, the input data can be a 3D tensor: . c t express t The state of the time step memory cell, h t Represents the output data for the previous time step. f t , i t ando t Represent the forget gate, input gate, and output gate respectively. The above process is repeated at each time step. In addition, σ Represents the sigmoid function, ” represents the element-by-element product. w represents different weight matrices, b Represents the bias vector. Figure 3 This is a schematic diagram of the storage unit structure of an LSTM in this specification.

[0062] BiLSTM processes the input sequence through two independent LSTM networks. One LSTM processes the sequence in forward order (from front to back), and the other LSTM processes the sequence in reverse order (from back to front). The forward LSTM output and the reverse LSTM output are combined to capture the contextual information of the sequence and improve the model's ability to understand global information. Therefore, BiLSTM is used to capture the time series data of air conditioning cooling loads at different locations and the time series change characteristics of various load impact data. Figure 4 This is a schematic diagram of a BiLSTM network architecture in this specification.

[0063] Furthermore, in one or more embodiments of the present invention, in order to enable the model to adaptively adjust the attention weight according to the temporal characteristics of the input sequence, thereby capturing the temporal characteristics of air-conditioning loads at different locations, the present invention designs a dynamic temporal attention mechanism (Dynamic Temporal Attention Mechanism) to more accurately capture the long-term and short-term dependencies between different time steps in the time series data of air-conditioning cooling loads at different locations and various types of load impact data by dynamically adjusting the attention weights.

[0064] As for the dynamic temporal attention mechanism, the attention mechanism is a method for dynamically selecting and aggregating important information when processing sequence data. The attention mechanism was originally introduced for neural machine translation, but it has now been widely used in various natural language processing, computer vision and other fields. Its core idea is to enable the model to pay attention to more relevant information by calculating the weighted sum of the input data, thereby improving the overall performance.

[0065] The core idea of ​​the dynamic temporal attention mechanism is to dynamically assign attention weights according to temporal features, so that the model can more flexibly learn and focus on time-related features in the data. During certain key time periods of the day (such as peak hours), the model can automatically assign higher attention weights to more accurately capture the characteristics of these time periods. Compared with the traditional fixed attention mechanism, the dynamic temporal attention mechanism can adapt to temporal changes and provide more sophisticated temporal feature modeling.

[0066] ① First, in order for the model to understand time information, the time features need to be encoded into a vector form, called time embedding. The time embedding vector can be generated using the following method:

[0067] Periodic Time Embeddings: Use sine and cosine functions to encode time information (hours, weeks, months) into a vector of fixed dimension. The formula is as follows: .

[0068] in t It's the time point. T is the time period (such as the total number of hours in a day, 24). When obtaining the time characteristics of air conditioning loads at different locations, in the design of basic time embedding, multiple embedding matrices are used to represent different time scales (hours, weeks, months), and their weighted summation is used to obtain the final time embedding vector. This method can help the model better capture the characteristics at different time scales and improve the modeling ability of time series data.

[0069] Define three time scales: Use the embedding matrix E h Representing hours, using the embedding matrix E d Representing the day of the week, using the embedding matrix E m Represents the month, and introduces a learnable weight parameter to perform weighted summation of embeddings of different time scales to form the final time embedding vector: .

[0070] in, α h , α d and α m is a learnable weight parameter that controls the contribution of each time scale in the final embedding. To ensure that the sum of the weights is 1, the weights are normalized using the Softmax function: .

[0071] Then calculate the weighted sum to get the final time embedding vector E final ( t ).

[0072] Use the time embedding vector to generate dynamic attention weights. Suppose the input sequence is X =[ x 1 , x 2 ,..., x n], the corresponding time embedding vector is T =[ t 1 , t 2 ,..., t n ], in order to generate temporal weights based on the temporal embedding vector, a neural network layer is used, with a weight matrix W and the bias term b : .

[0073] in, W and b is a parameter that needs to be learned, and the Softmax function ensures that the sum of the weights at all time points is 1.

[0074] ②Introducing time context attention: At the same time, a new attention mechanism is introduced, which not only weights each time point, but also considers the relationship between time points. This is very effective for processing long-term dependencies in time series data. A self-attention mechanism is used to capture the relationship between time points and construct query, key, and value matrices: .

[0075] in W Q , W K , W V is a learnable weight matrix.

[0076] ③ Calculate the similarity between time points: Use the dot product to calculate the similarity between each time point, and use the Softmax function to normalize it to get the attention weight between time points: .

[0077] in, α ij Indicates time point i About time point j The attention weight, d k is the dimension of the key vector, used to scale the dot product result.

[0078] ④ Combine the temporal context information to generate the final temporal attention weight: The temporal context attention result at each time point α ij The corresponding value vector V j Weighted summation to generate contextual representation of time points C i : .

[0079] ⑤ Calculate the final attention output: Combine the temporal context representation with the original dynamic temporal attention weights w i Combined, we get the final output at each time point: .

[0080] Here C i Contains temporal context information and uses dynamic weights w i To adjust its contribution to the final output, so that the final attention output can make full use of the temporal dependency and contextual information in the time series.

[0081] In order to more effectively capture the features of different time periods, an adaptive time window is introduced to dynamically adjust the time range of the attention mechanism. A learnable parameter is introduced to dynamically adjust the window range: .

[0082] in, f is a learnable function that can learn the optimal window size through back-propagation.

[0083] For the cross-attention mechanism, the cross-attention mechanism is used to combine the information of one sequence with the information of another sequence. It is very useful in tasks that require a relationship between two different sequences, such as machine translation tasks in the encoder-decoder architecture. The specific implementation steps of the cross-attention mechanism are:

[0084] ①Assume there is an input source sequence X =[ x 1 , x 2 ,..., x n ], target sequence Y =[ y 1 , y 2 ,..., y n ],in n is the length of the sequence.

[0085] ②Query matrix Q is generated by the target sequence, and the key matrix K Sum Matrix V Generated from the source sequence, where Q = YW Q , K = XWK , V = XW V ,in W Q , W K , W V is the learned weight matrix.

[0086] ③ Use the dot product to calculate the similarity between the query and the key, and then apply the scaling factor and the softmax function. The formula is as follows: .

[0087] ④The output is represented as a weighted sum of value vectors: .

[0088] For the crested porcupine optimization algorithm, the crested porcupine optimizer is a novel metaheuristic algorithm, and the CPO algorithm uses four different protection mechanisms: vision, sound, smell, and physical attack. The first and second defense strategies (vision and sound) reflect the exploration behavior of CPO, while the third and fourth defense strategies (smell and physical attack) reflect the exploitation behavior of CPO.

[0089] CPO proposes a new method: cyclic population reduction technique (CPR) to maintain the diversity of the population while speeding up convergence. This strategy simulates that not all CPs activate the defense mechanism, but only those CPs that are threatened will activate the defense mechanism. During the optimization process, this method takes some CPs out of the population to speed up convergence and puts them back into the population to improve diversity and avoid falling into local minima: .

[0090] in, T is the variable that determines the number of cycles, t is the current function evaluation, T max is the maximum number of function evaluations, % represents the remainder or modulo operator, N CPOmin is the minimum number of individuals in the newly generated population, so the population size cannot be less than N CPOmin .

[0091] In the exploration phase, according to the CP's defensive behavior, when the predator is far away, the CP has two defense strategies, namely the visual strategy and the sound strategy. These strategies refer to investigating different areas and are committed to global exploration: , .

[0092] in, is the evaluation function t The optimal solution of is a vector generated between the current CP and a randomly selected CP from the population, used to represent the predator at the iteration t The position at τ 1 is a random number based on a normal distribution, τ 2 is a random value in the interval [0,1], r is [1, N ] is a random number between ].

[0093] CP uses vocal methods to make noise and threaten predators. When a predator approaches, the CP's voice becomes louder: .

[0094] In the exploitation phase, according to the CP's defensive behavior, when a predator approaches, the CP has two defense strategies, namely the scent attack strategy and the physical attack strategy. These strategies refer to the exploitation of potential areas and the devotion to (local) exploitation search. In the scent attack phase, the CP will secrete a foul odor that spreads in the surrounding area to prevent predators from approaching it: .

[0095] U 1 Used to simulate three possible situations in this strategy:

[0096] (1) When U 1 If it is equal to 0, CP will stop the scent diffusion because the predator will stop moving due to fear of CP, so the distance between the predator and CP remains unchanged;

[0097] (2) When U 1 When it is equal to 1, CP will emit a significant smell because the predator is nearby;

[0098] (3) When U 1 is a combination of 0 and 1, the predator keeps a safe distance from the CP, so there is no need to release its scent widely.

[0099] in, r 3 is [1, N ], δ is a parameter used to control the search direction and is defined using the following formula: .

[0100] Iteration t Time i The location of an individual, γ t It's the defense factor. τ 3 is a random value in the interval [0,1], is the odor diffusion factor. It is as follows: , .

[0101] in, Represents iteration t Time i The objective function value of each individual, is a small value that avoids division by zero, rand is a variable containing a randomly generated value between 0 and 1, N CPO is the overall size, t is the number of current iterations, t max is the maximum number of iterations.

[0102] In the physical attack phase, when the predator is very close to it, the CP will attack it with short and thick feathers: .

[0103] in, α is the convergence speed factor, τ 4 and τ 5 is a random value in the interval [0,1], Is the impact i The average force of the CP of a predator.

[0104] In the CPO-BiSCN-BiLSTM-Attention model, a parallel deep neural network technology consisting of BiSCN, BiLSTM and Attention is used to capture the temporal variation characteristics and spatial distribution characteristics of sample data. The topological structure of the proposed BiSCN-BiLSTM-Attention method is shown in the figure. Figure 5 As shown, Figure 5The figure is a schematic diagram of a BiSCN-BiLSTM-Attention topology structure in the present invention. Specifically, the BiSCN on the left and the BiLSTM plus dynamic time attention layer on the right run in parallel to capture the spatial distribution characteristics and time series change characteristics of the air conditioning cooling load time series data and various load impact data at different locations, and then the two outputs are entered into the cross attention layer for information fusion, and the CPO algorithm is used to optimize the structure and parameters of the proposed BiSCN-BiLSTM-Attention neural network model, so as to accurately predict the radiant cooling load.

[0105] Among them, the server can first mark the cooling load time series data of air conditioners at different positions in the building air conditioning system at the historical target time after the sample data corresponds to the historical time as the real load data. Then the sample data is input into the prediction model to obtain the predicted load data, and the prediction model is trained for multiple rounds with the minimum deviation between the predicted load data and the real load data as the optimization goal. In each round of training, the crown porcupine optimization algorithm is used to optimize the number of convolution kernels of the bidirectional spatial convolution network, the number of hidden units and dropout rate of the bidirectional long short-term memory network, and the hidden layer size of the dynamic time attention module and the cross attention module.

[0106] Among them, the crowned porcupine optimization algorithm uses the prediction accuracy of the preset test samples as the fitness function; the fitness function can be the predicted root mean square error of the test samples weighted in different time periods, and the weights of different time periods represent the importance of the prediction accuracy of different time periods.

[0107] The BiSCN part includes a convolution layer, a layer normalization layer, a Dropout layer, a ReLu activation layer and a jump connection part, and the BiLSTM part includes a BiLSTM layer. The BiLSTM part is connected to the improved dynamic time attention layer, and the obtained output and the output of BiSCN enter the cross attention layer, and the final output enters the fully connected layer, and the regression output layer generates the final output. In order to improve the performance of radiative cooling load prediction, the present invention adopts the CPO algorithm to optimize the structure and parameters of the proposed neural network. CPO is used to simultaneously optimize the number of convolution kernels of BiSCN, the number of hidden units and Dropout rate of BiLSTM, the dynamic adjustment coefficient of the dynamic time attention mechanism, and the hidden layer size of the dynamic time attention mechanism and the cross attention mechanism. In the CPO algorithm, each particle in the particle swarm (i.e., a possible set of hyperparameters) can include 6 variables, representing different configurations of network hyperparameters.

[0108] In one or more embodiments of the present invention, the search space of hyperparameters may specifically include: the number of convolution kernels of BiSCN, the number of hidden units and Dropout rate of BiLSTM, the dynamic adjustment coefficient of the dynamic time attention mechanism, and the hidden layer size of the dynamic time attention mechanism and the cross attention mechanism, which are set to random numbers in the intervals [16, 128], [32, 256], [0.1, 0.5], [0.1, 0.5], [32, 512] and [32, 512] respectively. The search space of hyperparameters is determined considering the trade-off between computing time and prediction efficiency, and they are defined as integers. The CPO algorithm calculates the fitness function of each particle based on the prediction accuracy of the preset test samples by constructing a prediction model using the selected hyperparameters, such as the piecewise weighted root mean square error (Weighted RMSE).

[0109] In the task of air conditioning load prediction, different time periods (such as peak period and non-peak period) may have different importance. If the prediction model performs poorly during the peak period, it may lead to greater economic losses or system instability. Therefore, in one or more embodiments of the present invention, different weights may be assigned to different time periods in the fitness function to make the prediction model perform better during the peak period. The RMSE formula of the piecewise weighted is: ,in, N is the total number of test samples, g i To assign weights, represents the forecast load data of the forecast model, y i Represents the real load data. Dynamic weighting can be used for weighting. The peak period weight is , the weight of off-peak period is 1.

[0110] Since the mask operation was introduced above, the loss function of the prediction model should also be improved. The mask matrix can be introduced M , , which is valid data. When it is represented as masked data, the new Masked MSE formula is: .

[0111] The hyperparameter optimization procedure of the proposed network is calculated as follows: , In the formula, Represents the preferred hyperparameters of the network.

[0112] The prediction model simultaneously utilizes the powerful spatial distribution feature extraction capability of BiSCN and the long-range dependency capture of BiLSTM. By introducing the temporal context self-attention mechanism, the model can not only focus on the characteristics of each time point itself, but also capture the relationship between time points. This is very effective for processing long-term dependencies in time series data. In addition, refined feature weighting is performed, so the proposed method can achieve more accurate and efficient predictions.

[0113] In summary, the CPO-BiSCN-BiLSTM-Attention framework aims to ensure real-time monitoring of radiant cooling load changes to accurately understand the cooling needs of different locations in the building under different time periods and environmental conditions. This helps to optimize the operation strategy of the air-conditioning system and avoid unnecessary energy consumption. Due to the nature of data-driven methods based on deep learning, a large amount of training data is required to accurately learn the degradation process. A large amount of training data will make the data-driven method very slow and unable to meet real-time requirements. This framework uses BiSCN and BiLSTM parallel operations to greatly accelerate the feature extraction process, which is more efficient than traditional recurrent neural networks (RNNs). The CPO algorithm aims to adjust the hyperparameters of the proposed framework to improve the load prediction accuracy. Therefore, although the CPO-BiSCN-BiLSTM-Attention model integrates multiple data-driven methods to improve the prediction accuracy of radiant cooling load, it does not reduce the calculation speed, so it still meets the characteristics of accurate and fast prediction. Figure 6 This is a flow chart of a CPO-BiSCN-BiLSTM-Attention model in the present invention.

[0114] In one or more embodiments of the present invention, the CPO-BiSCN-BiLSTM-Attention prediction model constructed by the present invention is completed in Matlab, the sample data is input into the prediction model, the BiSCN activation function is set to ReLu, the number of convolutional layers is 4, the number of BiLSTM layers is 1, the number of heads of the dynamic time attention mechanism and the cross attention mechanism are 3 and 8 respectively, the learning rate is 0.001, the batch size is 32, the number of training rounds is 50, the optimizer can select the Adam optimizer, the dynamic time attention mechanism sets multiple time scales (including days, weeks, months), and uses the following combined dynamic time: hours: [0, 23], days: [0, 6], months: [1, 12], and the standard He initialization method is used when initializing the attention weights, which helps training convergence, and the initial size of the adaptive time window is set to a smaller value of 10. The remaining parameters include the number of convolution kernels of BiSCN, the number of hidden units and Dropout rate of BiLSTM, the dynamic adjustment coefficient of the dynamic time attention mechanism, and the hidden layer size of the dynamic time attention mechanism and the cross attention mechanism. The CPO optimization algorithm is used to optimize the loss function. After finding the optimal hyperparameter combination, the neural network introduced in this study uses the mean square error (MSE) as the loss function during the training process. Due to the introduction of masking operations, the loss function must also be improved. Since the Adam optimizer can provide superior prediction performance, this study uses it to optimize the loss function of CPO-BiSCN-BiLSTM-Attention during the model training stage. CPO optimization algorithm parameter settings: N CPO = 30, T max =100, α = 0.2, T f =0.8, T =2, N CPOmin = 0.8 N CPO .

[0115] After completing the training of the prediction model, the root mean square error (RMSE), the correlation coefficient square R 2 , Mean absolute error (MAE) and mean absolute percentage error (MAPE) were used as evaluation criteria to analyze the prediction performance. R 2 The larger the value is and the closer it is to 1, the better the prediction effect of the model is; the smaller the RMSE, MAE, and MAPE values ​​are, the higher the accuracy of the model prediction is; the specific formula is as follows: , , , .

[0116] in, Represents the average value of the real load data.

[0117] Finally, the trained building air-conditioning system load prediction model can be applied. In practical applications, the data to be predicted can be input into the building air-conditioning system load prediction model, and the spatial distribution characteristics of the time series data of air-conditioning cooling load at different locations and the influence relationship characteristics of various load influence data and the time series data of air-conditioning cooling load at different locations are extracted through the bidirectional spatial convolutional network. The time series change characteristics of the time series data of air-conditioning cooling load at different locations and various load influence data in the data to be predicted are extracted through the long short-term memory neural network. The time series change characteristics, spatial distribution characteristics and influence relationship characteristics are fused through the cross-attention module to obtain the fused characteristics, so as to determine the cooling load of each air conditioner in the building air-conditioning system at the target time according to the fused characteristics.

[0118] Among them, after extracting the time series change characteristics of the air-conditioning cooling load time series data and various load influence data at different locations in the data to be predicted through the long short-term memory neural network, the dynamic time attention module can also be used to perform time weighting on the corresponding features at different times in the time series change characteristics, so as to obtain the fused features by fusing the time-weighted time series change characteristics, spatial distribution characteristics and influence relationship characteristics through the cross attention module.

[0119] based on Figure 1 The load prediction method for a building air-conditioning system shown in the figure first obtains the historical time series data of the cooling load of air conditioners at different positions in the building air-conditioning system and the load impact data of the corresponding historical time period as sample data, and then obtains a bidirectional spatial convolutional network based on the BiTCN network improvement, connects the bidirectional long short-term memory network and the dynamic time attention module in series to form a parallel structure with the bidirectional spatial convolutional network, and connects the cross attention module in series after the parallel structure to construct a prediction model, wherein the dynamic time attention module is obtained by connecting a self-attention module in series after multiple parallel time embedding modules of different time scales; then, the prediction model is trained with sample data, and the hyperparameters are optimized with the crown porcupine optimization algorithm to obtain a load prediction model for the building air-conditioning system; finally, the load prediction model for the building air-conditioning system is used to predict the cooling load of each air conditioner in the building air-conditioning system at the target time based on the data to be predicted.

[0120] The dynamic temporal attention mechanism enables the model to dynamically assign weights based on load characteristics in different time periods. This flexibility improves the model's sensitivity to important temporal features and enables it to better adapt to load changes in different scenarios. It solves the problem that existing models assume that the input contributions of all time steps are the same, allowing the model to more effectively identify and respond to load changes in different time periods (such as peak and trough periods). For example, at noon in summer, air conditioning load tends to rise sharply, and the model can automatically strengthen the prediction weight of this period, thereby improving prediction accuracy.

[0121] The model design of the present invention allows it to perform effective feature selection and learning in high-dimensional data, especially in the case of multiple meteorological parameters and multi-region load data. By combining spatial convolution and time series change feature extraction, the model can integrate complex input features. It solves the shortcomings of existing simple models in processing multi-dimensional data and ensures that the model can still work effectively under high-dimensional input conditions. For complex commercial buildings, it can comprehensively consider multiple meteorological parameters (such as temperature, humidity, wind speed) and load data to provide more accurate predictions.

[0122] The present invention first runs two powerful deep learning methods, BiSCN and BiLSTM, in parallel. BiSCN efficiently captures spatial distribution features through convolution operations, and BiLSTM effectively processes long-term dependencies and contextual information through gating mechanisms and memory cells. A single BiSCN or BiLSTM model may have limitations in feature extraction. BiSCN mainly focuses on spatial distribution features, but is not as good as BiLSTM in capturing long-term dependencies. Although a single BiLSTM is good at processing long-term dependencies, it is not as good as BiSCN in processing spatial distribution features. Combining the two can capture features in time series at different levels and perspectives, and comprehensively utilize the advantages of convolution and recurrent neural networks to improve the comprehensiveness and accuracy of feature extraction.

[0123] The parallel computing of the present invention utilizes the parallel processing advantages of BiSCN to speed up the feature extraction. At the same time, the recursive computing of BiLSTM can also be optimized in the parallel structure, which improves the computing efficiency of the model as a whole. Therefore, the parallel structure designed by the present invention can effectively improve the computing efficiency of the model, so that the model structure is complex but still has a faster computing speed.

[0124] The cross-attention mechanism of the present invention provides an interpretable weight distribution in the feature fusion process, making the decision-making process of the model more transparent and interpretable, which helps to understand the working principle of the model and the direction of improvement. It is more interpretable than the ordinary black box model.

[0125] The BiSCN designed by the present invention can make full use of the feature extraction ability of the convolutional network. When extracting the spatial distribution characteristics of air-conditioning load data at different locations, the forward 2D convolution will reduce the spatial resolution of the feature map and may cause the loss of details. Therefore, the reverse transposed convolution operation is added to restore the low-resolution feature map to a higher resolution. The transposed convolution helps to restore the global details lost by the forward convolution. Through this process, the details can be reconstructed to generate a clearer and higher-quality output. By using the bidirectional processing mode of BiTCN, the advantages of the original bidirectional processing sequence time characteristics are utilized. There may be spatial dependencies between different locations of air-conditioning load data. The bidirectional convolution operation can capture these dependencies more comprehensively, improve the model's modeling ability for spatial distribution characteristics, and the reverse operation of the transposed convolution can compensate for the boundary, making the feature extraction at the boundary more complete. Therefore, the designed BiSCN performs well in extracting spatial distribution characteristics, enables the model to understand and express spatial distribution characteristics more comprehensively, improves the expression ability of features and the adaptability of the model, and is a method for effectively capturing the spatial distribution feature extraction of air-conditioning loads at different locations.

[0126] In order to improve the expressive power of temporal features, the improved dynamic temporal attention mechanism adopts multi-scale temporal embedding, which not only uses a single time scale, but also combines multiple time scales (hours, weeks, months). This multi-scale embedding method enables the model to learn features at different time scales at the same time. At the same time, the temporal context attention mechanism is introduced to better capture the relationship between time points. This mechanism not only focuses on the features of each time point, but also considers the interdependence and relationship between time points, which is particularly useful when processing time series data. Finally, in order to more effectively capture the features of different time periods, an adaptive time window is introduced to dynamically adjust the time range of the attention mechanism. By dynamically adjusting the attention weight, the improved mechanism can more accurately capture the long-term and short-term dependencies in the sequence, which is particularly important for processing sequence data with complex time dependencies. Dynamic temporal attention can dynamically adjust the attention weight according to the actual changes in the data, instead of using a fixed attention mechanism. This adaptability enables the model to automatically adjust its attention focus at different time steps, thereby better handling changing time patterns.

[0127] In the process of model training, the present invention adopts an optimization algorithm (CPO) to automatically adjust network parameters, including the number of convolution kernels of BiSCN, the number of hidden units and Dropout rate of BiLSTM, the dynamic adjustment coefficient of dynamic time attention mechanism, the hidden layer size of dynamic time attention mechanism and cross attention mechanism, etc. This automatic optimization mechanism not only reduces the workload of manual parameter adjustment, but also adjusts the model structure according to data characteristics, so as to achieve better performance in different time series prediction tasks.

[0128] The method proposed in the present invention significantly improves the performance of radiant cooling load prediction, accurately makes load predictions for different locations in a building, optimizes building energy consumption management, helps the building management system to accurately control the operating time and intensity of the refrigeration system, and reduces unnecessary energy consumption. By predicting the air-conditioning load requirements at different locations in advance, the system can reduce the refrigeration power when the load demand is low to avoid energy waste. Provide real-time data support to help the management system dynamically adjust the refrigeration strategy according to actual conditions and optimize energy use. The present invention also aims to reduce energy waste and operating costs, while supporting environmental protection and sustainable development, reducing carbon emissions, and contributing to environmental protection and addressing climate change. Efficient energy use is an important part of achieving green buildings and sustainable development. It has a significant promoting effect on building energy conservation.

[0129] When applying the building air conditioning system load prediction method provided by the present invention, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0130] The above is a method for predicting the load of a building air conditioning system provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for predicting the load of a building air conditioning system, such as Figure 7 shown.

[0131] Figure 7 A schematic diagram of a load prediction device for a building air conditioning system provided by the present invention includes:

[0132] An acquisition module 201 is used to acquire historical time series data of cooling loads of air conditioners at different locations in a building air conditioning system and load impact data of corresponding historical periods as sample data;

[0133] A construction module 202 is used to convert the extended causal convolution layer in the bidirectional temporal convolutional network into a 2D convolution to obtain a bidirectional spatial convolutional network, and connect a cross attention module in series after the parallel structure formed by the bidirectional spatial convolutional network and the bidirectional long short-term memory network to construct a prediction model;

[0134] A training module 203 is used to train a prediction model using a crown porcupine optimization algorithm according to sample data to obtain a load prediction model for a building air conditioning system;

[0135] The prediction module 204 is used to input the data to be predicted into the load prediction model of the building air-conditioning system, extract the spatial distribution characteristics of the time series data of the air-conditioning cooling load at different locations and the influence relationship characteristics of various load influence data and the time series data of the air-conditioning cooling load at different locations through a bidirectional spatial convolution network, extract the time series change characteristics of the time series data of the air-conditioning cooling load at different locations and various load influence data in the data to be predicted through a long short-term memory neural network, and fuse the time series change characteristics, spatial distribution characteristics and influence relationship characteristics through a cross-attention module to obtain a fusion feature, so as to determine the cooling load of each air-conditioner in the building air-conditioning system at the target time according to the fusion feature.

[0136] The specific definition of the load prediction device for the building air conditioning system can be found in the definition of the load prediction method for the building air conditioning system mentioned above, which will not be repeated here. Each module in the above-mentioned load prediction device for the building air conditioning system can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0137] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A load prediction method for building air conditioning systems is provided.

[0138] The present invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 A load prediction method for building air conditioning systems is provided.

[0139] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0140] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A method for predicting load of a building air conditioning system, characterized in that: include: Obtain the cooling load historical time series data of air conditioners at different locations in the building air conditioning system and the load impact data of the corresponding historical period as sample data; The extended causal convolution layer in the bidirectional temporal convolutional network is converted into a 2D convolution to obtain a bidirectional spatial convolutional network. The cross-attention module is connected in series after the parallel structure formed by the bidirectional spatial convolutional network and the bidirectional long short-term memory network to build a prediction model. The cooling load time series data of air conditioners at different positions in the building air conditioning system at the historical target moment after the sample data corresponds to the historical moment are marked as the real load data; the sample data are input into the prediction model to obtain the predicted load data, and the prediction model is trained for multiple rounds with the minimum deviation between the predicted load data and the real load data as the optimization goal. In each round of training, the crown porcupine optimization algorithm is used to optimize the number of convolution kernels of the bidirectional spatial convolutional network, the number of hidden units and the dropout rate of the bidirectional long short-term memory network, and the hidden layer size of the dynamic time attention module and the cross attention module to obtain the building air conditioning system load prediction model; The data to be predicted is input into the load prediction model of the building air conditioning system. The spatial distribution characteristics of the time series data of the air conditioning cooling load at different locations and the influence relationship characteristics of various load influence data and the time series data of the air conditioning cooling load at different locations are extracted through the bidirectional spatial convolution network. The time series change characteristics of the time series data of the air conditioning cooling load at different locations and various load influence data in the data to be predicted are extracted through the long short-term memory neural network. The time series change characteristics, spatial distribution characteristics and influence relationship characteristics are fused through the cross attention module to obtain the fusion characteristics, so as to determine the cooling load of each air conditioner in the building air conditioning system at the target time according to the fusion characteristics; The load impact data includes: at least one of outdoor temperature, indoor temperature, outdoor humidity, indoor humidity, solar radiation intensity, wind speed, holiday attribute information, real-time number of people in each area of ​​the building, and water supply temperature and return water temperature of the air conditioning system; The crested porcupine optimization algorithm uses the prediction accuracy of the preset test samples as the fitness function; the fitness function is the prediction root mean square error of the test samples weighted in different time periods, and the weights of different time periods represent the importance of the prediction accuracy of different time periods; The influence relationship feature represents the influence relationship between various load influence data at different locations and the air-conditioning cooling load time series data at different locations.

2. The building air conditioning system load prediction method according to claim 1, characterized in that: The obtaining of cooling load historical time series data of air conditioners at different locations in the building air conditioning system and load impact data of corresponding historical periods specifically includes: Obtain historical time series data of cooling load of air conditioners at different locations in the building air conditioning system; Obtain multiple types of load impact data for the corresponding historical period, including outdoor temperature, indoor temperature, outdoor humidity, indoor humidity, solar radiation intensity, wind speed, holiday attribute information, real-time number of people in the building, and supply and return water temperatures of the air conditioning system; For the air conditioner at each location, the cosine similarity between the historical time series data of the cooling load of the air conditioner at that location and various types of load impact data is determined, so as to screen various types of load impact data according to the cosine similarity and the preset similarity threshold, and determine one or more types of load impact data for load forecasting at that location.

3. The building air conditioning system load prediction method according to claim 2, characterized in that: The 2D convolution is a 2D masked convolution; The extracting of the influence relationship characteristics of various load impact data and air conditioning cooling load time series data at different locations through the bidirectional spatial convolutional network specifically includes: For the air conditioner at each location, mask the data other than the load impact data used for load forecasting at the location in various types of load impact data through 2D mask convolution; Extract the influence relationship characteristics of various types of load impact data after masking and the time series data of air-conditioning cooling load at different locations.

4. The building air conditioning system load prediction method according to claim 1, characterized in that: The parallel structure includes two branches, the first branch includes a bidirectional spatial convolutional network, and the second branch includes a bidirectional long short-term memory network and a dynamic time attention module connected in series; wherein the dynamic time attention module is obtained by connecting a plurality of parallel time embedding modules of different time scales in series with the self-attention module; The long short-term memory neural network is used to extract the time series change characteristics of the air conditioning cooling load time series data at different locations in the data to be predicted and the time series change characteristics, spatial distribution characteristics and influence relationship characteristics are fused by the cross attention module to obtain the fusion characteristics, which specifically include: The time series change characteristics of air conditioning cooling load time series data and various load impact data at different locations in the data to be predicted are extracted through long short-term memory neural network; The dynamic time attention module is used to perform time weighting on the features corresponding to different times in the time series change features; The cross-attention module is used to fuse the time-weighted temporal change features, spatial distribution features and influence relationship features to obtain the fused features.

5. The building air conditioning system load prediction method according to claim 1, characterized in that: The load impact data includes: at least one of outdoor temperature, indoor temperature, outdoor humidity, indoor humidity, solar radiation intensity, wind speed, holiday attribute information, real-time number of people in each area of ​​the building, and water supply temperature and return water temperature of the air-conditioning system.

6. The building air conditioning system load prediction method according to claim 5, characterized in that: The method further comprises: In the process of training the prediction model based on sample data, the random forest algorithm is used to perform sensitivity tests on various types of load impact data in the sample data, and the redundant data types in various types of load impact data are determined and eliminated according to the preset sensitivity threshold.

7. A load prediction device for a building air conditioning system, characterized in that: include: An acquisition module is used to acquire historical time series data of cooling loads of air conditioners at different locations in the building air conditioning system and load impact data of corresponding historical periods as sample data; A construction module is used to convert the extended causal convolution layer in the bidirectional temporal convolutional network into a 2D convolution to obtain a bidirectional spatial convolutional network, and connect a cross-attention module to the parallel structure formed by the bidirectional spatial convolutional network and the bidirectional long short-term memory network to build a prediction model; A training module is used to label the cooling load time series data of air conditioners at different positions in the building air conditioning system at the historical target moment after the historical moment corresponding to the sample data as real load data; The sample data is input into the prediction model to obtain the predicted load data. The prediction model is trained for multiple rounds with the minimum deviation between the predicted load data and the real load data as the optimization goal. In each round of training, the crown porcupine optimization algorithm is used to optimize the number of convolution kernels of the bidirectional spatial convolutional network, the number of hidden units and the dropout rate of the bidirectional long short-term memory network, and the hidden layer size of the dynamic time attention module and the cross attention module to obtain the load prediction model of the building air conditioning system. The prediction module is used to input the data to be predicted into the load prediction model of the building air conditioning system, extract the spatial distribution characteristics of the time series data of the air conditioning cooling load at different locations and the influence relationship characteristics of various load influence data and the time series data of the air conditioning cooling load at different locations through a bidirectional spatial convolution network, extract the time series change characteristics of the time series data of the air conditioning cooling load at different locations and various load influence data in the data to be predicted through a long short-term memory neural network, and fuse the time series change characteristics, spatial distribution characteristics and influence relationship characteristics through a cross attention module to obtain a fusion feature, so as to determine the cooling load of each air conditioner in the building air conditioning system at the target time according to the fusion feature; The load impact data includes: at least one of outdoor temperature, indoor temperature, outdoor humidity, indoor humidity, solar radiation intensity, wind speed, holiday attribute information, real-time number of people in each area of ​​the building, and water supply temperature and return water temperature of the air conditioning system; The crested porcupine optimization algorithm uses the prediction accuracy of the preset test samples as the fitness function; the fitness function is the prediction root mean square error of the test samples weighted in different time periods, and the weights of different time periods represent the importance of the prediction accuracy of different time periods; The influence relationship feature represents the influence relationship between various load influence data at different locations and the air-conditioning cooling load time series data at different locations.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

Images