Cooling and heating load control method, system and equipment of temperature control equipment and storage medium

Abstract

The invention belongs to the technical field of cooling and heating load control of equipment, and provides a cooling and heating load control method and system of temperature control equipment, the equipment and a storage medium, and the method comprises the steps: obtaining historical load data of the temperature control equipment; preprocessing the historical load data to obtain preprocessed data; analyzing the preprocessed data by adopting a target model to obtain a predicted value; and performing unit start-stop control on the temperature control equipment based on the predicted value. According to the method, the target model can effectively capture the cold and hot load change trend in a long time sequence through a self-attention mechanism, the accuracy of load prediction is remarkably improved, and particularly for a complex and multi-dimensional load change scene, compared with an existing simple trend analysis and regression model, the accuracy of load prediction is greatly improved. According to the method, the load peak or valley can be estimated in advance, and the problem that a traditional method cannot accurately cope with long-time-period changes is avoided.

Description

Technical Field

[0001] This invention belongs to the field of equipment cooling and heating load control technology, and specifically relates to cooling and heating load control methods, systems, equipment and storage media for temperature control equipment. Background Technology

[0002] Cooling and heating load forecasting refers to predicting the cooling and heating needs of a building or facility for a future period based on historical data, environmental conditions, and building characteristics. It is a crucial task in building energy management, especially in large buildings, commercial spaces, and industrial facilities, where the energy efficiency of heating, ventilation, and air conditioning (HVAC) systems directly impacts energy consumption and operating costs. Therefore, accurate cooling and heating load forecasting can significantly optimize energy scheduling and usage, reduce unnecessary energy consumption, and improve user comfort.

[0003] Traditional methods for forecasting heating and cooling loads primarily rely on statistical models, such as ARIMA (Autoregressive Integral Moving Average) and linear regression, as well as machine learning-based models, such as support vector machines and LSTM (Long Short-Term Memory). While these methods perform well in short-term load forecasting, they have significant limitations when dealing with load forecasting over longer time series. 1. Effective for short-term forecasts, ineffective for long-term forecasts: Most models are good at forecasting loads in the short term (such as within one day), but their accuracy drops rapidly for forecasts over many days, months or even years.

[0004] 2. High computational complexity: Some advanced neural network models, such as LSTM, can capture non-linear relationships in time series, but they have high computational complexity when dealing with long sequence data and are prone to gradient vanishing or exploding problems.

[0005] 3. Data sparsity and noise issues: Hot and cold load data are often sparse and noisy, which makes it difficult for traditional models to effectively handle long-term forecasts, especially when the data is incomplete or highly volatile.

[0006] Furthermore, the Transformer model has achieved great success in the field of natural language processing. Transformer-based models have been introduced into time series forecasting to address the problem of long sequence dependencies. However, the standard Transformer also faces several challenges when applied to time series forecasting: 1. High computational complexity: The self-attention mechanism of the standard Transformer has high computational complexity. When processing long-term sequences, the computational cost and memory requirements are too high, making it difficult to apply to large-scale data scenarios.

[0007] 2. Slow decoding speed: The stepwise decoding mechanism of the traditional Transformer is inefficient in long sequence prediction, resulting in a significant decrease in inference speed.

[0008] In addition, current transformer-based methods require a large amount of data, and can only obtain predicted heating and cooling loads, but cannot achieve direct control of the unit. Summary of the Invention

[0009] To address the problems in the background art, this invention proposes a method, system, device, and storage medium for controlling the cold and heat loads of temperature control equipment.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions: A method for controlling the heating and cooling loads of a temperature control device includes the following steps: Obtain historical load data of temperature control equipment; Historical load data is preprocessed to obtain preprocessed data; The target model is used to analyze the preprocessed data and obtain predicted values, including: using a self-attention mechanism to reduce the computational complexity of the preprocessed data and outputting a self-attention weight matrix; convolving the self-attention weight matrix through a self-attention distillation mechanism to obtain the time series; and analyzing the time series through a generative decoder to obtain predicted values. The unit starts and stops based on the predicted values.

[0011] Preferably, a self-attention mechanism is used to reduce the complexity of preprocessed data, including the following steps: Transform the preprocessed data into a query matrix, a value matrix, and a key matrix; By introducing sparsity conditions, query matrices, value matrices, and key matrices that satisfy the sparsity conditions are obtained, thereby reducing the computational complexity of preprocessed data. Input the query matrix, value matrix, and key matrix that satisfy the sparsity condition into the self-attention mechanism formula to obtain the self-attention weights.

[0012] Preferably, the formula for the self-attention mechanism is: ; In the formula, Q Represents the query matrix; V Represents a value matrix; K Represents the key matrix; Indicates about Q , K and V Self-attention weights; d k The dimension representing the key; The sparsity condition is satisfied as follows: ; In the formula, L K The set of all keys, i.e., the length of the key sequence; q i This represents the i-th query; k j This represents the j-th key; M ( q i ,K )express q i The corresponding sparsity measure.

[0013] Preferably, the self-attention weight matrix is ​​compressed using a self-attention distillation mechanism, including: ; In the formula, represents the output sequence of the self-attention weights for the (j+1)th iteration; Conv1d represents a one-dimensional convolution operation, ELU represents an activation function, and MaxPool represents a max pooling operation; Indicates the first 20 M ( q i ,K ) corresponding q i The matrix formed; K Represents the key matrix; V Represents the value matrix; Attention ( ) indicates about Self-attention weights.

[0014] Preferably, the predicted value is obtained by analyzing the time series using a generative decoder, including: ;

[0015] In the formula, This represents the "token sequence" that serves as the starting point for prediction. Placeholders indicating the sequence to be predicted; Y pred This indicates that the decoder calculates and generates predicted load values ​​for several future time steps through a self-attention mechanism.

[0016] Preferably, the unit start-up and shutdown control based on the predicted value includes the following steps: Set the start-stop threshold for each unit in the temperature control equipment; Set the number of units operating at the initial moment during the temperature control equipment's operation; Calculate the total supply capacity of the units during operation; Start-up and shutdown decisions are made based on total supply capacity, forecast values, and start-up / shutdown thresholds. The start-stop sequence of the temperature control equipment is updated based on the start-stop judgment results.

[0017] Preferably, the start / stop determination is based on the total supply capacity, forecast value, and start / stop threshold, including the following steps: like and In this case, additional generating units will be started, and the supply capacity of the existing generating units will be updated. express t+p Predicted value at time, express t+p An estimate of the unit's supply capacity during operation at any given time; T on Indicates the threshold for starting the generator unit; like ,and If necessary, some generating units will be shut down, and the supply capacity of the operating units will be updated. T off This indicates the threshold for unit shutdown.

[0018] Preferably, updating the unit start-up and shutdown sequence of the temperature control equipment based on the start-up and shutdown determination results includes the following steps: Record the number of units in operation at the time of startup detection; The start-up and shutdown sequence of the units is constructed based on the number of units in operation at the corresponding time of the start-up determination.

[0019] A temperature control system for heating and cooling loads, comprising: The data acquisition unit is used to acquire historical load data of the temperature control equipment; The preprocessing unit is used to preprocess historical load data to obtain preprocessed data; The prediction unit, used to analyze preprocessed data using the target model and obtain predicted values, includes: The sparse module is used to reduce the computational complexity of preprocessed data by employing a self-attention mechanism and outputs a self-attention weight matrix. The distillation module is used to convolve the self-attention weight matrix through a self-attention distillation mechanism to obtain the time series. The prediction module is used to analyze the time series using a generative decoder to obtain predicted values; The control unit is used to control the start-up and shutdown of the temperature control equipment based on predicted values.

[0020] Preferably, the sparse module includes: The transformation submodule is used to convert preprocessed data into query matrices, value matrices, and key matrices; The sparse submodule is used to introduce sparse conditions to obtain query matrices, value matrices and key matrices that satisfy the sparse conditions, thereby reducing the computational complexity of preprocessed data. The computation submodule is used to input the query matrix, value matrix, and key matrix that satisfy the sparsity condition into the self-attention mechanism formula to obtain the self-attention weights.

[0021] Preferably, the control unit includes: The first setting module is used to set the start-stop threshold for each unit in the temperature control equipment; The second setting module is used to set the number of units operating at the initial time during the operation of the temperature control equipment; The calculation module is used to calculate the total supply capacity of the unit during operation; The judgment module is used to make start-up and shutdown decisions based on the total supply capacity, the predicted value, and the start-up and shutdown thresholds. The update module is used to update the unit start-stop sequence of the temperature control equipment based on the start-stop judgment result.

[0022] An apparatus comprising: Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the above-mentioned method for controlling the hot and cold loads of a temperature control device.

[0023] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for controlling the hot and cold loads of a temperature control device.

[0024] The beneficial effects of this invention are: 1. The target model of the method of the present invention can effectively capture the trend of cold and hot load changes in long time series through the self-attention mechanism, which significantly improves the accuracy of load forecasting. Especially for complex and multi-dimensional load change scenarios, compared with the existing simple trend analysis and regression models, the above method can predict the load peak or trough in advance, avoiding the problem that traditional methods cannot accurately cope with long-term changes. 2. The method of the present invention uses sparse conditions to filter the processing of the self-attention mechanism, thereby reducing the processing required by the self-attention mechanism and reducing the computational difficulty; 3. This invention uses the predicted values ​​of the target model to judge and adjust the start-up and shutdown status of temperature control equipment, making the prediction of cooling and heating loads more efficient in practical applications and adapting to the needs of large-scale data and real-time prediction. It provides more accurate and faster prediction results for energy management systems, thereby helping to reduce energy waste, optimize equipment control, and ultimately improve the overall performance of the system.

[0025] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description and the drawings. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A flowchart of a method for controlling the cold and heat loads of a temperature control device according to the present invention is shown; Figure 2 A structural diagram of the Informer model of the present invention is shown; Figure 3 The diagram illustrates the principle of load forecasting and start / stop strategy distribution based on the Informer model of the present invention. Figure 4 A structural diagram of a temperature control device's heating and cooling load control system according to the present invention is shown. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] like Figure 1 The image shows a method for controlling the heating and cooling loads of a temperature control device, comprising the following steps: S1: Obtain historical load data of the temperature control equipment.

[0030] S2: Preprocess the historical load data to obtain preprocessed data.

[0031] S3: The target model is used to analyze the preprocessed data and obtain the predicted value, including: S301: The self-attention mechanism (ProbSparse) is used to reduce the computational complexity of the preprocessed data and output the self-attention weight matrix; S302: The self-attention weight matrix is ​​convolved through the self-attention distillation mechanism to obtain the time series; S303: The time series is analyzed through the generative decoder to obtain the predicted value.

[0032] S4: Control the start-up and shutdown of the unit based on the predicted value.

[0033] It should be noted that the target model used in the above method is the Informer model. This model, by introducing a sparse self-attention mechanism, can effectively capture the trend of cold and hot load changes over long periods, significantly improving the accuracy of load forecasting, especially for complex and multidimensional load change scenarios. Compared with existing simple trend analysis and regression models, the above method can predict load peaks or troughs in advance, avoiding the problem that traditional methods cannot accurately cope with long-term changes.

[0034] In S1, the temperature control equipment can be a building central air conditioning system, which can be equipped with temperature sensors, humidity sensors, and flow sensors. These sensors are deployed at key locations in the air conditioning units, ducts, and indoors and outdoors. The sensors can collect data at daily and hourly intervals and store it as monitoring data for cooling and heating loads. This data can provide support for subsequent model processing. In addition, this historical data mainly includes: the unit's supply and return water temperatures, flow rates, start / stop status, set temperatures, energy consumption, meteorological data (outdoor temperature, humidity, solar radiation intensity), indoor environmental data (indoor temperature, humidity, carbon dioxide concentration), and the building's occupancy density, etc.

[0035] In S2, the preprocessing process involves cleaning the collected data to remove noisy data, outliers, and missing values, ensuring data purity. Secondly, data standardization processing, such as normalization, is performed to give the data consistent feature representation, enabling the Informer model to better learn data features. In addition, interpolation processing can be used for some missing data to ensure data integrity and facilitate the Informer model's subsequent understanding of data patterns.

[0036] like Figure 2 The diagram shown is a block diagram of the Informer model, which includes an encoder and a decoder. Generally, the data input to the Informer model is historical time series data. The encoder can contain timestamp information, which is mapped to a high-dimensional space through special encoding. The timestamp data is then concatenated with historical time-series data. This data is then transmitted to a sparse self-attention module, which significantly reduces computation by sampling the top-weighted samples using a self-attention mechanism on the high-dimensional matrix as a representation of the original data. The encoder contains multiple sparse self-attention modules, whose output data sizes are not uniform, exhibiting pooling characteristics. The final output of the encoder is then input to the decoder (the input is...). The multi-head self-attention module in the representation performs cross-attention operations with the masked sparse self-attention module, and the output is passed through a fully connected layer to obtain the final one-step or multi-step prediction output. The decoder input needs to add empty sequences to the end of the original input sequence for padding, and if there are timestamps, the timestamp features of the prediction also need to be taken into account.

[0037] In S3, S301 may specifically include the following steps: S3011: Convert the preprocessed data into a query matrix, value matrix, and key matrix. S3012: Introduce sparsity conditions to obtain query matrices, value matrices, and key matrices that satisfy the sparsity conditions, thereby reducing the computational complexity of the preprocessed data. S3013: Input the query matrix, value matrix, and key matrix that satisfy the sparsity conditions into the self-attention mechanism formula to obtain the self-attention weights.

[0038] It should be noted that for the Informer model, the input data includes the query matrix Q, the value matrix V, and the key matrix K. Therefore, in S3011, the preprocessed data needs to be converted into the corresponding Q, V, and K. Then, in S3012, to reduce computational cost, the self-attention mechanism introduces the query sparsity assumption. This assumption assumes that most query points contribute little to the computation, so only important query points are selected for calculation, i.e., data that satisfies the sparsity condition is chosen. Finally, in S3013, the important query points are input into the self-attention mechanism formula to obtain the final self-attention weight matrix.

[0039] The formula for the self-attention mechanism is as follows: (1) In the formula, Q Represents the query matrix; V Represents a value matrix; K Represents the key matrix; Indicates about Q , K and V Self-attention weights; additionally... ,in These are the sequence lengths of the query, key, and value, respectively. These are the dimensions of the key and the value, respectively. It is a scaling factor used to prevent the gradient from vanishing due to an excessively large dot product result; It is the input to the Informer model; W Q , W K and W V This represents a learnable weight matrix used to weight the input. Mapped to the query space, the query matrix is ​​obtained through matrix multiplication.Q Key matrix K Sum matrix V ; express Q The space of the real matrix; express K The space of the real matrix; express V The space of the real matrix in which it resides.

[0040] The sparsity condition satisfies the following formula: (2) In the formula, L K The number of keys, i.e., the length of the key sequence; q i This represents the i-th query; k j This represents the j-th key; M ( qi,K )express q i The corresponding sparsity measure is expressed as: q i and K The higher the correlation of the matrix, the stronger the correlation.

[0041] Then, from the matrix Q Extract the top 20 most relevant ones q i , build Matrix. Execute the following formula: (3) Under the sparsity condition, only the top few that satisfy the sparsity condition are selected. The calculation is performed on each query point, reducing the computational complexity from the original... Reduced to .

[0042] In S302, self-attention distillation effectively reduces redundant information in the input data by extracting and filtering attention information layer by layer, enabling the model to handle extremely long input sequences. The distillation mechanism formula is as follows: (4) In the formula, This represents the output sequence of the self-attention weights for the (j+1)th iteration; Conv1d represents a one-dimensional convolution operation; ELU represents an activation function; and MaxPool represents a max pooling operation (typically with a stride of 2, effectively halving the length). Indicates the first 20 M ( q i,K ) corresponding q i The matrix formed; K Represents the key matrix; V Represents the value matrix; Attention ( ) indicates about Self-attention weights.

[0043] Formula (3) extracts important features in the time dimension through 1D convolution, and then halves the length of the time series through max pooling. Finally, through a multi-layer distillation mechanism, the length of the input data is compressed from L to L / 2 and L / 4, reducing the computational overhead and memory usage of the model.

[0044] In S303, the time series is analyzed using a generative decoder to obtain predicted values, including: (5) In the formula, This represents the "token sequence" that serves as the starting point for prediction. Placeholders representing the sequence to be predicted. Y pred This indicates that the decoder calculates and generates predicted load values ​​for several future time steps through a self-attention mechanism.

[0045] It should be noted that the input of the decoder It consists of two parts: one part is a known "token sequence" that serves as the starting point for prediction. (For example, a known load and related characteristic sequence before the prediction begins), the other part is a placeholder for the sequence to be predicted. (Usually filled with 0, but includes timestamp information of the target sequence).

[0046] It's important to note that generative decoders aim to generate a complete prediction result with a single forward pass, avoiding the cumulative error problem common in stepwise decoding. For the Transformer, the output is generated by first outputting the first result, then the second based on the first, and so on. This is slow and inaccurate. Looking at the overall architecture diagram, we can see that the decoder consists of two parts: the first part is the encoder's output, and the second part is the decoder input after embedding, i.e., the latter half of the input is masked with 0s. This method obtains multiple outputs from a single input, significantly improving computational efficiency.

[0047] S4 specifically includes the following steps: S401: Set the start-stop threshold for each unit in the temperature control equipment. Specifically, in order to determine the start-stop of the units, it is necessary to set the start-stop threshold (start-up threshold) for each unit. Shutdown threshold ), assuming there is Two units with the same power output have the same heating / cooling capacity. Q total Furthermore, the coefficient of performance (COP) of the unit is within the range of 3 to 4.

[0048] S402: Set the number of units in operation at the initial moment during the operation of the temperature control equipment, denoted as S(t).

[0049] S403: Calculate the total supply capacity of the unit during operation, denoted as... Q total ( t ).

[0050] S404: Start-up and shutdown decisions are made based on total supply capacity, forecast values, and start-up / shutdown thresholds, specifically including: like and In this case, additional generating units will be started, and the supply capacity of the existing generating units will be updated. ,in express t+p Predicted value at time, express t+p An estimate of the unit's supply capacity during operation at any given time; T on Indicates the threshold for starting the generator unit; express t+P The total supply capacity of the generating units at any given moment during operation; like ,and If necessary, some generating units will be shut down, and the supply capacity of the operating units will be updated. ,in T off This indicates the threshold for unit shutdown.

[0051] S405: Update the unit start-stop sequence of the temperature control equipment based on the start-stop judgment result, including: S4051: Record the number of units in operation at the time corresponding to the start-up judgment.

[0052] S4052: Construct the start-up and shutdown sequence of units based on the number of units in operation at the corresponding time of the start-up judgment.

[0053] It should be noted that the above solution intelligently adjusts the unit start-up and shutdown times based on load forecasting results, effectively reducing unnecessary frequent start-ups and shutdowns. In particular, by predicting peak loads in advance, the solution can pre-start the units before the load reaches its peak, avoiding full-load operation during peak periods, reducing system energy consumption and extending equipment lifespan. Simultaneously, intelligent adjustment of start-up and shutdown times significantly reduces equipment maintenance costs. Furthermore, accurate load forecasting and optimized start-up and shutdown control enable the central air conditioning system to operate with minimal energy consumption while maintaining indoor comfort. This energy-saving effect is particularly significant in large-scale buildings, saving users substantial energy costs and reducing environmental impact, aligning with green building and sustainable development requirements.

[0054] like Figure 3 As shown, after making a start / stop decision, the Informer model provides start / stop suggestions through the cloud, then forms a strategy and sends it to the air conditioning, heating, and ventilation (HVAC) system for execution. Finally, the data from the execution is also input into the Informer model for processing, forming a closed loop.

[0055] A temperature control system for hot and cold loads includes a data acquisition unit, a preprocessing unit, a prediction unit, and a control unit.

[0056] The acquisition unit is used to acquire historical load data of the temperature control equipment; the preprocessing unit is used to preprocess the historical load data to obtain preprocessed data; the prediction unit is used to analyze the preprocessed data using a target model to obtain predicted values; and the control unit is used to control the start-up and shutdown of the temperature control equipment based on the predicted values.

[0057] Furthermore, the prediction unit includes a sparse module, a distillation module, and a prediction module. The sparse module is used to reduce the computational complexity of the preprocessed data using a self-attention mechanism and outputs a self-attention weight matrix; the distillation module is used to convolve the self-attention weight matrix through a self-attention distillation mechanism to obtain the time series; and the prediction module is used to analyze the time series through a generative decoder to obtain the predicted value.

[0058] In addition, the sparse module includes a transformation submodule, a sparse submodule, and a computation submodule. The transformation submodule is used to convert the preprocessed data into query rectangles, value matrices, and key matrices; the sparse submodule is used to introduce sparse conditions to obtain query rectangles, value matrices, and key matrices that satisfy the sparse conditions, thereby reducing the computational complexity of the preprocessed data; the computation submodule is used to input the query matrix, value matrix, and key matrix that satisfy the sparse conditions into the self-attention mechanism formula to obtain the self-attention weights.

[0059] Furthermore, the control unit includes a first setting module, a second setting module, a calculation module, a judgment module, and an update module. The first setting module is used to set the start-stop threshold for each unit in the temperature control equipment; the second setting module is used to set the number of units operating at the initial moment during operation of the temperature control equipment; the calculation module is used to calculate the total supply capacity of the operating units; the judgment module is used to make start-stop judgments based on the total supply capacity, predicted values, and start-stop thresholds; and the update module is used to update the unit start-stop sequence of the temperature control equipment according to the results of the start-stop judgments.

[0060] An apparatus comprising: Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the above-mentioned method for controlling the hot and cold loads of a temperature control device.

[0061] It should be noted that the memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device.

[0062] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0063] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method for controlling the hot and cold loads of a temperature control device.

[0064] It should be noted that the computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus. The computer-readable storage medium carries one or more programs, which, when executed, implement a method for controlling the cold and heat loads of a temperature control device according to an embodiment of the present invention.

[0065] According to embodiments of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium, such as: a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0066] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for controlling the heating and cooling loads of a temperature control device, characterized in that, Includes the following steps: Obtain historical load data of temperature control equipment; Historical load data is preprocessed to obtain preprocessed data; The target model is used to analyze the preprocessed data and obtain predicted values, including: using a self-attention mechanism to reduce the computational complexity of the preprocessed data and outputting a self-attention weight matrix; convolving the self-attention weight matrix through a self-attention distillation mechanism to obtain the time series; and analyzing the time series through a generative decoder to obtain predicted values. The unit starts and stops based on the predicted values.

2. The method for controlling the heating and cooling loads of a temperature control device according to claim 1, characterized in that, Using a self-attention mechanism to reduce the complexity of preprocessed data includes the following steps: Transform the preprocessed data into a query matrix, a value matrix, and a key matrix; By introducing sparsity conditions, query matrices, value matrices, and key matrices that satisfy the sparsity conditions are obtained, thereby reducing the computational complexity of preprocessed data. Input the query matrix, value matrix, and key matrix that satisfy the sparsity condition into the self-attention mechanism formula to obtain the self-attention weights.

3. The method for controlling the heating and cooling loads of a temperature control device according to claim 2, characterized in that, The formula for the self-attention mechanism is: ； In the formula, Q Represents the query matrix; V Represents a value matrix; K Represents the key matrix; Indicates about Q , K and V Self-attention weights; d k The dimension representing the key; The sparsity condition is satisfied as follows: ； In the formula, L K The set of all keys, i.e., the length of the key sequence; q i This represents the i-th query; k j This represents the j-th key; M ( q i ,K )express q i The corresponding sparsity measure.

4. The method for controlling the heating and cooling loads of a temperature control device according to claim 1, characterized in that, Compressing the self-attention weight matrix using a self-attention distillation mechanism includes: ； In the formula, represents the output sequence of the self-attention weights for the (j+1)th iteration; Conv1d represents a one-dimensional convolution operation, ELU represents an activation function, and MaxPool represents a max pooling operation; Indicates the first 20 M ( q i ,K ) corresponding q i The matrix formed; K Represents the key matrix; V Represents the value matrix; Attention ( ) indicates about Self-attention weights.

5. The method for controlling the heating and cooling loads of a temperature control device according to claim 1, characterized in that, The predicted values ​​are obtained by analyzing the time series using a generative decoder, including: ； In the formula, This represents the "token sequence" that serves as the starting point for prediction. Placeholders indicating the sequence to be predicted; Y pred This indicates that the decoder calculates and generates predicted load values ​​for several future time steps through a self-attention mechanism.

6. The method for controlling the heating and cooling loads of a temperature control device according to claim 1, characterized in that, The unit start-up and shutdown control based on predicted values ​​for temperature control equipment includes the following steps: Set the start-stop threshold for each unit in the temperature control equipment; Set the number of units operating at the initial moment during the temperature control equipment's operation; Calculate the total supply capacity of the units during operation; Start-up and shutdown decisions are made based on total supply capacity, forecast values, and start-up / shutdown thresholds. The start-stop sequence of the temperature control equipment is updated based on the start-stop judgment results.

7. The method for controlling the heating and cooling loads of a temperature control device according to claim 6, characterized in that, The start / stop determination is based on total supply capacity, forecast values, and start / stop thresholds, including the following steps: like and In this case, additional generating units will be started, and the supply capacity of the existing generating units will be updated. express t+p Predicted value at time, express t+p An estimate of the unit's supply capacity during operation at any given time; T on Indicates the threshold for starting the generator unit; like ,and If necessary, some generating units will be shut down, and the supply capacity of the operating units will be updated. T off This indicates the threshold for unit shutdown.

8. The method for controlling the heating and cooling loads of a temperature control device according to claim 6, characterized in that, The start-up and shutdown sequence of the temperature control equipment is updated based on the start-up and shutdown determination results, including the following steps: Record the number of units in operation at the time of startup detection; The start-up and shutdown sequence of the units is constructed based on the number of units in operation at the corresponding time of the start-up determination.

9. A temperature control system for both heating and cooling loads, characterized in that, include: The data acquisition unit is used to acquire historical load data of the temperature control equipment; The preprocessing unit is used to preprocess historical load data to obtain preprocessed data; The prediction unit, used to analyze preprocessed data using the target model and obtain predicted values, includes: The sparse module is used to reduce the computational complexity of preprocessed data by employing a self-attention mechanism and outputs a self-attention weight matrix. The distillation module is used to convolve the self-attention weight matrix through a self-attention distillation mechanism to obtain the time series. The prediction module is used to analyze the time series using a generative decoder to obtain predicted values; The control unit is used to control the start-up and shutdown of the temperature control equipment based on predicted values.

10. A temperature control system for hot and cold loads according to claim 9, characterized in that, The sparse module includes: The transformation submodule is used to convert preprocessed data into query matrices, value matrices, and key matrices; The sparse submodule is used to introduce sparse conditions to obtain query matrices, value matrices and key matrices that satisfy the sparse conditions, thereby reducing the computational complexity of preprocessed data. The computation submodule is used to input the query matrix, value matrix, and key matrix that satisfy the sparsity condition into the self-attention mechanism formula to obtain the self-attention weights.

11. The cold and heat load control system for a temperature control device according to claim 9, characterized in that, The control unit includes: The first setting module is used to set the start-stop threshold for each unit in the temperature control equipment; The second setting module is used to set the number of units operating at the initial time during the operation of the temperature control equipment; The calculation module is used to calculate the total supply capacity of the unit during operation; The judgment module is used to make start-up and shutdown decisions based on the total supply capacity, the predicted value, and the start-up and shutdown thresholds. The update module is used to update the unit start-stop sequence of the temperature control equipment based on the start-stop judgment result.

12. A device, characterized in that, include: Memory, used to store computer programs; When a processor executes a program stored in a memory, it implements the method for controlling the hot and cold loads of a temperature control device as described in any one of claims 1-8.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for controlling the cold and hot loads of a temperature control device as described in any one of claims 1-8.

Images