Method and system for load prediction of constant temperature and humidity air conditioning system based on transfer learning

Abstract

The method and system for load prediction of constant temperature and humidity air conditioning system based on transfer learning comprise the following steps: obtaining historical factory environment parameter sequences and corresponding loads under different external environment parameters of different factories, extracting feature data, clustering all feature data, taking the historical factory environment parameter sequences and corresponding loads of the same cluster as a training set to train a load prediction model; obtaining a current factory environment parameter sequence, calculating the sequence similarity of the current factory environment parameter sequence and the cluster centers of all clusters, taking the load prediction model with the maximum sequence similarity as the benchmark load prediction model; constructing a target domain model with the same network structure as the benchmark load prediction model, performing transfer learning on the target domain model, inputting the current factory environment parameter sequence into the target domain model, and outputting the predicted load; and judging the air conditioner start-stop state of the factory at the next moment to perform thermal inertia correction. The present application solves the bottleneck that a single model is difficult to adapt to multiple factories and variable environments.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent control of air conditioning systems, and more particularly relates to a constant temperature and humidity air conditioning system load prediction method and system based on transfer learning. BACKGROUND

[0002] Constant temperature and humidity air conditioning systems are widely used in data centers, power production workshops and other scenarios with high requirements for environmental temperature and humidity accuracy. The core goal of such systems is to maintain constant environmental conditions through real-time regulation of temperature and humidity. However, the load of air conditioning systems is affected by various nonlinear factors, including external environmental conditions (such as temperature and humidity changes), indoor thermal and humidity loads (equipment operation and personnel activity), and system operating conditions. In addition, the energy consumption of constant temperature and humidity air conditioning systems accounts for 60%-70% of the total energy consumption of these scenarios, and the load prediction and efficient control of these systems are directly related to the energy saving effect and sustainable operation of the system.

[0003] Traditional air conditioning system load prediction methods are usually based on simple linear models or single time series analysis methods, which are difficult to cope with the nonlinear dynamic changes and fluctuations of the load under complex environments. In addition, due to the scarcity or incompleteness of historical data in actual engineering, the cold start problem of new systems further reduces the generalization ability and prediction accuracy of traditional methods. Therefore, there is an urgent need for an intelligent load prediction method that combines physical laws and data-driven methods, which can not only accurately capture complex load change trends, but also automatically adapt to different operating scenarios and provide practical decision support for subsequent energy-saving control. SUMMARY

[0004] To solve the above problems, the purpose of the present application is to provide a constant temperature and humidity air conditioning system load prediction method and system based on transfer learning, which realizes accurate control and energy efficiency maximization of constant temperature and humidity workshop air conditioning systems.

[0005] The application adopts the following technical solutions.

[0006] The first aspect of the application proposes a constant temperature and humidity air conditioning system load prediction method based on transfer learning, including the following contents:

[0007] Obtain historical factory environmental parameter sequences and corresponding loads under different factory external environmental parameters, the historical factory environmental parameter sequences including factory temperature data sequences and factory humidity data sequences within a set period; extract feature data from each historical factory environmental parameter sequence, cluster all feature data, and use the same cluster of historical factory environmental parameter sequences and their corresponding loads as a training set to train a load prediction model;

[0008] obtaining a current factory environment parameter sequence, calculating a sequence similarity between the current factory environment parameter sequence and a historical factory environment parameter sequence corresponding to a cluster center of all clusters, and taking a load prediction model corresponding to a cluster center with the largest sequence similarity as a benchmark load prediction model;

[0009] taking the benchmark load prediction model as a source domain model, constructing a target domain model with the same network structure as the source domain model, performing transfer learning on the target domain model, inputting the current factory environment parameter sequence into the target domain model, and outputting a predicted load;

[0010] judging an air conditioner start-stop state of the factory at a next time according to the predicted load, and correcting the predicted load for shutdown or startup thermal inertia according to the air conditioner start-stop state.

[0011] Preferably, the feature data includes mean values, variances, maximum values, minimum values and quartile ranges of the factory temperature data sequence and the factory humidity data sequence, and average moisture content and average wet enthalpy value; the average moisture content is an average value of all moisture contents in a set period; and the average wet enthalpy value is an average value of all moisture contents in a set period.

[0012] Preferably, the clustering of all feature data is specifically as follows:

[0013] each feature data corresponding to each factory environment parameter sequence is normalized to serve as a sample, a maximum cluster number is set, for each cluster number from 2 to the maximum cluster number, K-medoids algorithm is used to cluster the samples, an average distance of each sample to all samples of other clusters is calculated, a smallest average distance is selected, a difference between the smallest average distance and an average distance of all samples in the same cluster is divided by a maximum value between the smallest average distance and the average distance of all samples in the same cluster to obtain a silhouette coefficient, the silhouette coefficients under all cluster numbers are calculated, when a difference between a silhouette coefficient of a cluster number and a silhouette coefficient of a previous cluster number exceeds a set coefficient threshold, the cluster number is taken as a final cluster number of clustering; if no difference between a silhouette coefficient of a cluster number and a silhouette coefficient of a previous cluster number exceeds a set coefficient threshold, a silhouette coefficient with a maximum value among all cluster numbers is taken as a final cluster number of clustering; if there are multiple cluster numbers whose differences between the silhouette coefficients and the silhouette coefficient of the previous cluster number exceed the set coefficient threshold, a silhouette coefficient with a maximum value among the multiple cluster numbers is taken as a final cluster number of clustering.

[0014] Preferably, the load prediction model is specifically as follows:

[0015] The load prediction model includes an input layer, a feature encoder and a prediction head.

[0016] The input layer inputs a current plant environment parameter sequence; the feature encoder is a bidirectional LSTM layer, the bidirectional LSTM layer outputs a sequence composed of hidden states of each time step, and the prediction head performs quantile regression output to output a predicted load; the quantile regression output is to set a plurality of quantiles greater than 0 and less than 1, each quantile has a different weight matrix and bias, and the sequence composed of hidden states of each time step output by the bidirectional LSTM layer is multiplied by the weight matrix corresponding to the quantile and added to the bias corresponding to the quantile to obtain a prediction result corresponding to the quantile, and the prediction results of all quantiles are weighted and summed with the corresponding quantile as the weight to obtain the final predicted load.

[0017] Preferably, the sequence similarity of the current plant environment parameter sequence and the historical plant environment parameter sequence corresponding to the cluster center of each cluster is calculated, specifically:

[0018] For the sequence of the current plant environment parameter sequence and the historical plant environment parameter sequence corresponding to the cluster center of each cluster, after aligning the two sequences in time, a subsequence of a set length is extracted from each of the two sequences, the slope, curvature and amplitude of each data point of the two subsequences are calculated, the slope, curvature and amplitude of the data point are taken as its coordinates, the slope is the next data point of the data point minus the previous data point of the data point, and then divided by twice the time step length between the data points, when the data point is the first data point of the subsequence, the data point is taken as the previous data point of the data point, when the data point is the last data point of the subsequence, the data point is taken as the next data point of the data point; the curvatures of all data points in the same subsequence are equal; the amplitude is the data point minus the average of the data points in the subsequence; the Euclidean distance of the coordinates of all data points of the same sequence number in the two subsequences is calculated and summed, the data points in the two subsequences are reordered until the sum of the Euclidean distances of the coordinates of all data points of the same sequence number in the two subsequences reaches a minimum, and the reciprocal of the minimum sum plus 1 is taken, which is the sequence similarity of the two sequences.

[0019] Preferably, the target domain model is subjected to transfer learning, specifically:

[0020] The network weight of the benchmark load prediction model is taken as the initial network weight of the target domain model, and the weight matrix and bias of each quantile in the prediction head are retrained until the transfer learning loss function reaches a minimum, and the transfer learning loss function is:

[0021]

[0022] Wherein, The quantile loss kernel function is The quantile is The true load value at time step t is the predicted value of the quantile for time step t, T is the total time step, is a regularization coefficient, , is the iteration number, u is the weight matrix and bias of the quantile for time step t when the iteration number is is the norm of .

[0023] Preferably, when the iteration number reaches a set number of times during transfer learning, the loss function of the transfer learning is always greater than a set loss function threshold, or the test set is respectively input into the benchmark load prediction model and the target domain model for prediction, and the predicted results are respectively input into the loss function of the load prediction model, and the loss function calculation result of the target domain model is greater than the loss function calculation result of the benchmark load prediction model, then the weights of each bidirectional LSTM layer in the target domain model are retrained, and the loss function of the training is the loss function of the load prediction model.

[0024] Preferably, the air conditioner start-stop state of the factory at the next time is determined according to the predicted load, and specifically:

[0025] The air conditioner start-stop state includes a start-up phase, a normal operation phase and a stop phase.

[0026] The average value of the load of a set number of time points before calculation is calculated, if the difference between the predicted load and the average value of the load of the set number of time points before calculation is greater than a set difference threshold, and the predicted load is greater than a set start-stop threshold, then it is in the start-up phase.

[0027] If the difference between the average value of the load of the set number of time points before calculation and the predicted load is greater than a set difference threshold, or the predicted load is less than or equal to a set start-stop threshold, then it is in the stop phase.

[0028] Otherwise, it is in the normal operation phase.

[0029] Preferably, the predicted load is corrected for shutdown or startup thermal inertia according to the air conditioner start-stop state, and specifically:

[0030] The internal envelope area of the factory and the factory volume are obtained, and the internal envelope area includes internal walls and floors; the corrected load calculation formula is:

[0031]

[0032] wherein, is the corrected load; is the predicted load; , are respectively a startup correction coefficient and a shutdown correction coefficient.​​ is a time step between two time instants; is a thermal inertia constant; is an internal envelope area; is a factory volume.

[0033] The second aspect of the present application proposes a constant temperature and humidity air conditioning system load prediction system based on the method of the first aspect of the present application, comprising a load prediction model construction module, a benchmark load prediction model acquisition module, a transfer learning module and a correction module, specifically:

[0034] The load prediction model construction module: obtains historical factory environment parameter sequences and corresponding loads under different factory external environment parameters, the historical factory environment parameter sequences including factory temperature data sequences and factory humidity data sequences in a set period; feature data is extracted for each historical factory environment parameter sequence, all feature data is clustered, and the historical factory environment parameter sequences and their corresponding loads in the same cluster are used as a training set to train a load prediction model;

[0035] The benchmark load prediction model acquisition module: obtains the current factory environment parameter sequence, calculates the sequence similarity of the current factory environment parameter sequence and the historical factory environment parameter sequences corresponding to the cluster centers of all clusters, and takes the load prediction model corresponding to the cluster center with the maximum sequence similarity as the benchmark load prediction model;

[0036] The transfer learning module: takes the benchmark load prediction model as a source domain model, constructs a target domain model with the same network structure as the source domain model, performs transfer learning on the target domain model, inputs the current factory environment parameter sequence into the target domain model, and outputs the predicted load;

[0037] The correction module: determines the air conditioner start-stop state of the next time instant according to the predicted load, and performs shutdown or startup thermal inertia correction on the predicted load according to the air conditioner start-stop state.

[0038] The present application has the advantages that, compared with the prior art, the engineering environment parameter data is clustered, a prediction model is independently trained for each cluster, the optimal benchmark model is matched based on sequence similarity, and the network structure is copied and transferred to quickly adapt to the current working condition of the target factory. This method significantly reduces the training cost in a data-scarce scenario, improves the prediction reliability by matching the sequence similarity, solves the bottleneck of a single model being difficult to adapt to multiple factories and variable environments, improves the prediction accuracy of the model in a specific scenario, predicts the air conditioner start-stop decision according to the load prediction result, and introduces a thermal inertia correction mechanism, which takes into account the building environmental factors of the factory, avoids energy consumption fluctuations and equipment wear and tear caused by frequent start-stop, and optimizes load prediction and subsequent control. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0040] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. The embodiments described in the present application are only a part of the embodiments of the present application, but not all the embodiments. All other embodiments obtained by those skilled in the art without creative labor based on the spirit of the present application shall fall within the protection scope of the present application.

[0041] The present application adopts the technical solutions as follows.

[0042] As shown in Figure 1 Embodiment 1 of the present application proposes a constant temperature and humidity air conditioning system load prediction method based on transfer learning, which includes the following contents:

[0043] obtain historical factory environment parameter sequences and corresponding loads under different external environment parameters of different factories, wherein the historical factory environment parameter sequences include factory temperature data sequences and factory humidity data sequences in a set period; extract feature data of each historical factory environment parameter sequence, cluster all the feature data, and take the historical factory environment parameter sequences in the same cluster and the corresponding loads as a training set to train a load prediction model;

[0044] obtain a current factory environment parameter sequence, calculate the sequence similarity of the current factory environment parameter sequence and the historical factory environment parameter sequences corresponding to the cluster centers of all clusters, and take the load prediction model corresponding to the cluster center with the largest sequence similarity as a reference load prediction model;

[0045] take the reference load prediction model as a source domain model, construct a target domain model with the same network structure as the source domain model, perform transfer learning on the target domain model, input the current factory environment parameter sequence into the target domain model, and output the predicted load;

[0046] determine the air conditioner start-stop state of the factory at the next moment according to the predicted load, and perform shutdown or startup heat inertia correction on the predicted load according to the air conditioner start-stop state.

[0047] Preferably, the feature data includes the average value, variance, maximum value, minimum value and quartile range of the factory temperature data sequence and the factory humidity data sequence, as well as the average moisture content and the average wet enthalpy value; the average moisture content is the average value of all moisture contents in a set period; and the average wet enthalpy value is the average value of all moisture contents in a set period.

[0048] Specifically, the wet content calculation formula is:

[0049]

[0050] wherein, is the air water vapor partial pressure, is the atmospheric pressure.

[0051] The wet enthalpy value The calculation formula is:

[0052]

[0053] wherein, is the dry-bulb temperature, is the wet content.

[0054] Preferably, the clustering of all feature data is specifically:

[0055] Each feature data corresponding to each factory environment parameter sequence is normalized as a sample, the maximum cluster number is set, for each cluster number from 2 to the maximum cluster number, each sample is clustered by using the K-medoids algorithm, and the average distance of each sample to all samples of other clusters is calculated, the smallest average distance is selected, the average distance of each sample to all samples of the same cluster is calculated, the difference between the smallest average distance and the average distance of all samples of the same cluster is divided by the maximum of the smallest average distance and the average distance of all samples of the same cluster as a silhouette coefficient, the silhouette coefficient under all cluster numbers is calculated, when the difference between the silhouette coefficient of a cluster number and the silhouette coefficient of the previous cluster number exceeds the set coefficient threshold, the cluster number is taken as the final clustering cluster number; if the difference between the silhouette coefficient of a cluster number and the silhouette coefficient of the previous cluster number does not exceed the set coefficient threshold, the maximum silhouette coefficient among all cluster numbers is taken as the final clustering cluster number, and if there are multiple cluster numbers whose difference between the silhouette coefficient and the silhouette coefficient of the previous cluster number exceeds the set coefficient threshold, the maximum silhouette coefficient among the multiple cluster numbers is taken as the final clustering cluster number.

[0056] Preferably, the load prediction model is specifically:

[0057] The load prediction model comprises an input layer, a feature encoder and a prediction head.

[0058] The input layer inputs a current factory environment parameter sequence; it should be noted that the current factory environment parameter sequence is a sequence composed of a current factory temperature data sequence and a current factory humidity data sequence; the feature encoder is a bidirectional LSTM layer, the bidirectional LSTM layer outputs a sequence composed of hidden states of each time step, the prediction head performs quantile regression output, and outputs a predicted load; the quantile regression output is to set a plurality of quantiles greater than 0 and less than 1, for each quantile, there is a different weight matrix and bias, and the sequence composed of hidden states of each time step output by the bidirectional LSTM layer is multiplied by the weight matrix corresponding to the quantile and added to the bias corresponding to the quantile to obtain a prediction result corresponding to the quantile, and the prediction results of all quantiles are weighted and summed with the corresponding quantile as the weight to obtain the final predicted load.

[0059] It should be noted that the bidirectional LSTM layer is specifically:

[0060]

[0061]

[0062] wherein, , are hidden states of the forward LSTM at time steps t, t-1 respectively; is a feature vector of the input sequence at time step t; , are a forward LSTM unit and a backward LSTM unit respectively; , are hidden states of the forward LSTM at time steps t, t-1 respectively; is a hidden state of the bidirectional LSTM layer at time step t, is to splice and

[0063] It should be noted that the quantile represents the probability that the load value is less than or equal to the predicted value, and the quantiles are set to 0.1, 0.5 and 0.9 respectively. In addition, in addition to the final predicted load obtained by weighted fusion, the confidence interval of each prediction can also be obtained, for example, in this embodiment, 80% of the loads are located in the interval. , are the final predicted loads with quantiles of 0.1 and 0.9 respectively.

[0064] Loss function of the load prediction model is:

[0065] . ​

[0066] wherein, quantile loss kernel function, is a quantile, is a true load value at time step t, is a predicted value of quantile at time step t, T is a total time step.

[0067] The embodiment preferably calculates the sequence similarity of the current plant environment parameter sequence and the sequence of historical plant environment parameter sequences corresponding to the cluster centers of all clusters, specifically:

[0068] For the sequence of the current plant environment parameter sequence and the historical plant environment parameter sequence corresponding to the cluster center of each cluster, after aligning the two sequences in time, extracting a subsequence of a set length in the two sequences respectively, calculating the slope, curvature and amplitude of each data point of the two subsequences, taking the slope, curvature and amplitude of the data point as its coordinates, the slope is the next data point of the data point minus the last data point of the data point, and then divided by twice the time step length between the data points, when the data point is the first data point of the subsequence, the data point is used as the last data point of the data point, when the data point is the last data point of the subsequence, the data point is used as the next data point of the data point; the curvature of all data points in the same subsequence is equal; the amplitude is the data point minus the mean of the subsequence data points; calculate the Euclidean distance of the coordinates of all data points of the same sequence number in the two subsequences and sum them up, reorder the data points in the two subsequences until the sum of the Euclidean distances of the coordinates of all data points of the same sequence number in the two subsequences reaches the minimum, add 1 to the minimum sum result and take the reciprocal, the reciprocal is the sequence similarity of the two sequences.

[0069] The embodiment preferably performs transfer learning on the target domain model, specifically:

[0070] Taking the network weight of the benchmark load prediction model as the initial network weight of the target domain model, retraining the weight matrix and bias of each quantile in the prediction head until the transfer learning loss function reaches the minimum, the transfer learning loss function is:

[0071]

[0072] wherein, quantile loss kernel function, is a quantile, is a true load value at time step t, is a predicted value of quantile at time step t, T is a total time step, is a regularization coefficient, , are the number of iterations, respectively.u time division quantile weight matrix and bias of the quantile is norm of

[0073] It should be noted that the iterative formula for retraining the weight matrix and bias of each quantile in the prediction head is:

[0074]

[0075] , weight matrix and bias of the quantile u +1, respectively; weight matrix and bias of the quantile , weight matrix and bias of the quantile of the source domain model, learning rate of the set quantile .

[0076] Preferably, in the transfer learning, if the number of iterations reaches the set number of times, the loss function of the transfer learning is always greater than the set loss function threshold, or the test set is input into the benchmark load prediction model and the target domain model for prediction, respectively, and the predicted results are input into the loss function of the load prediction model, and the loss function calculation result of the target domain model is greater than the loss function calculation result of the benchmark load prediction model, then each weight of the bidirectional LSTM layer in the target domain model is retrained, and the loss function of the training is the loss function of the load prediction model.

[0077] Preferably, the air conditioner start-stop state of the factory at the next moment is determined according to the predicted load, and specifically:

[0078] The air conditioner start-stop state includes a start-up phase, a normal operation phase, and a stop phase.

[0079] The average value of the load at a set number of time points before is calculated, if the difference between the predicted load and the average value of the load at the set number of time points before is greater than a set difference threshold, and the predicted load is greater than a set start-stop threshold, then it is in the start-up phase.

[0080] If the difference between the average value of the load at the set number of time points before and the predicted load is greater than the set difference threshold, or the predicted load is less than or equal to the set start-stop threshold, then it is in the stop phase.

[0081] Otherwise, it is in the normal operation phase.

[0082] Preferably, the predicted load is corrected for shutdown or startup thermal inertia according to the air conditioner start-stop state, and specifically:

[0083] obtain the internal envelope area and the factory volume of the factory, the internal envelope including internal walls, floors, and the corrected load calculation formula is:

[0084]

[0085] wherein, is the corrected load; is the predicted load; , are respectively a start-up correction coefficient and a shutdown correction coefficient set, and specifically, the embodiment , are respectively set to 0.6 and 0.7; is a time step between two time points; is a thermal inertia constant, ranging from 0.5 to 2.0; is the internal envelope area; is the factory volume.

[0086] Embodiment 2 of the present application proposes a constant temperature and humidity air conditioning system load prediction system based on the method described in Embodiment 1 of the present application, including a load prediction model construction module, a benchmark load prediction model acquisition module, a transfer learning module, and a correction module, specifically:

[0087] The load prediction model construction module: obtains historical factory environment parameter sequences and corresponding loads under different factory external environment parameters, the historical factory environment parameter sequences including factory temperature data sequences and factory humidity data sequences within a set period; feature data is extracted for each historical factory environment parameter sequence, all feature data is clustered, and the historical factory environment parameter sequences and their corresponding loads in the same cluster are used as a training set to train a load prediction model;

[0088] The benchmark load prediction model acquisition module: obtains a current factory environment parameter sequence, calculates the sequence similarity of the current factory environment parameter sequence and the historical factory environment parameter sequences corresponding to the cluster centers of all clusters, and takes the load prediction model corresponding to the cluster center with the maximum sequence similarity as the benchmark load prediction model;

[0089] The transfer learning module: takes the benchmark load prediction model as a source domain model, constructs a target domain model with the same network structure as the source domain model, performs transfer learning on the target domain model, inputs the current factory environment parameter sequence into the target domain model, and outputs the predicted load;

[0090] The correction module: determines the air conditioning start-stop state of the factory at the next time point according to the predicted load, and performs thermal inertia correction on the predicted load according to the air conditioning start-stop state.

[0091] The present disclosure can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present disclosure.

[0092] It should be noted that the above-mentioned embodiments are only used to illustrate the technical solutions of the present application, not to limit it. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or equivalent replaced without departing from the spirit and scope of the present application. Any modification or equivalent replacement should be covered within the protection scope of the claims of the present application.

Claims

1. A load prediction method for a constant temperature and humidity air conditioning system based on transfer learning, characterized in that, The application comprises the following contents: obtaining historical factory environment parameter sequences and corresponding loads under different factory external environment parameters, wherein the historical factory environment parameter sequences comprise factory temperature data sequences and factory humidity data sequences within a set period; extracting feature data of each historical factory environment parameter sequence, clustering all feature data, taking the historical factory environment parameter sequences of the same cluster and the corresponding loads as a training set to train a load prediction model; the load prediction model comprises an input layer, a feature encoder and a prediction head; the input layer inputs a current factory environment parameter sequence; the feature encoder is a bidirectional LSTM layer, which outputs a sequence composed of hidden states of each time step; the prediction head performs quantile regression output to output a predicted load; the quantile regression output is a plurality of quantiles greater than 0 and less than 1, each quantile has a different weight matrix and bias, and the sequence composed of hidden states of each time step output by the bidirectional LSTM layer is multiplied by the weight matrix corresponding to the quantile and added to the bias corresponding to the quantile to obtain a prediction result corresponding to the quantile; the prediction results of all quantiles are weighted and summed with the corresponding quantiles as weights to obtain a final predicted load; obtaining a current factory environment parameter sequence, calculating the sequence similarity between the current factory environment parameter sequence and the historical factory environment parameter sequences corresponding to the cluster centers of all clusters, specifically: for the sequence of the current factory environment parameter sequence and the historical factory environment parameter sequence corresponding to the cluster center of each cluster, aligning the two sequences in time, respectively extracting subsequences of a set length in the two sequences, calculating the slope, curvature and amplitude of each data point in the two subsequences, taking the slope, curvature and amplitude of the data point as its coordinates, the slope is the next data point minus the previous data point of the data point, divided by twice the time step length between the data points, when the data point is the first data point of the subsequence, the data point is used as the previous data point of the data point, when the data point is the last data point of the subsequence, the data point is used as the next data point of the data point; the curvatures of all data points in the same subsequence are equal; the amplitude is the data point minus the average of the subsequence data points; calculating the Euclidean distance of the coordinates of all data points with the same sequence number in the two subsequences and summing them up, reordering the data points in the two subsequences until the sum of the Euclidean distances of the coordinates of all data points with the same sequence number in the two subsequences reaches a minimum, adding 1 to the minimum sum and taking the reciprocal, which is the sequence similarity of the two sequences; taking the load prediction model corresponding to the cluster center with the maximum sequence similarity as the reference load prediction model; taking the reference load prediction model as a source domain model, constructing a target domain model with the same network structure as the source domain model, and performing transfer learning on the target domain model, specifically: taking the network weights of the reference load prediction model as the initial network weights of the target domain model, retraining the weight matrix and bias of each quantile in the prediction head until the transfer learning loss function reaches a minimum, the transfer learning loss function is: wherein, quantile loss kernel function, is a quantile, is a true load value at time step t, is a predicted value of the quantile at time step t, T is the total time step, is a regularization coefficient, , are the weight matrix and bias of the quantile u at iteration i, respectively, is the norm of .​ Inputting the current factory environment parameter sequence into the target domain model outputs the predicted load; According to the predicted load, the air conditioner start-stop state of the factory at the next time is determined, and the predicted load is corrected for shutdown or startup thermal inertia according to the air conditioner start-stop state.

2. The constant temperature and humidity air conditioning system load prediction method based on transfer learning according to claim 1, characterized in that: The feature data includes the average value, variance, maximum value, minimum value and quartile range of the factory temperature data sequence and the factory humidity data sequence, and the average moisture content and the average wet enthalpy value; the average moisture content is the average value of all moisture contents in a set period; and the average wet enthalpy value is the average value of all moisture contents in a set period.

3. The constant temperature and humidity air conditioning system load prediction method based on transfer learning according to claim 1, characterized in that: The clustering of all feature data is specifically as follows: After normalizing each feature data corresponding to each factory environment parameter sequence, the normalized data is taken as a sample, the maximum cluster number is set, for each cluster number from 2 to the maximum cluster number, the K-medoids algorithm is used to cluster each sample, the average distance of each sample to all samples in other clusters is calculated, the smallest average distance is selected, the average distance of each sample to all samples in the same cluster is calculated, the difference between the smallest average distance and the average distance of all samples in the same cluster is divided by the maximum value between the smallest average distance and the average distance of all samples in the same cluster to obtain the silhouette coefficient, the silhouette coefficients of all cluster numbers are calculated, when the difference between the silhouette coefficient of a cluster number and the silhouette coefficient of the previous cluster number exceeds the set coefficient threshold, the cluster number is taken as the final clustering cluster number; if the difference between the silhouette coefficients of all cluster numbers and the silhouette coefficient of the previous cluster number does not exceed the set coefficient threshold, the maximum silhouette coefficient among all cluster numbers is taken as the final clustering cluster number; if there are multiple cluster numbers whose difference between the silhouette coefficient and the silhouette coefficient of the previous cluster number exceeds the set coefficient threshold, the maximum silhouette coefficient among the multiple cluster numbers is taken as the final clustering cluster number.

4. The constant temperature and humidity air conditioning system load prediction method based on transfer learning according to claim 1, characterized in that: When the number of iterations reaches the set number of iterations during transfer learning, the loss function of transfer learning is always greater than the set loss function threshold, or the test set is input into the benchmark load prediction model and the target domain model for prediction, the predicted results are input into the loss function of the load prediction model, and the loss function calculation result of the target domain model is greater than the loss function calculation result of the benchmark load prediction model, then the weights of the bidirectional LSTM layer in the target domain model are retrained, and the loss function of the training is the loss function of the load prediction model.

5. The constant temperature and humidity air conditioning system load prediction method based on transfer learning according to claim 1, characterized in that: The air conditioner start-stop state of the factory at the next time is determined according to the predicted load, and is specifically as follows: The air conditioner start-stop state includes a starting phase, a normal running phase and a stopping phase. The average value of the load at a preset number of time points before calculation is set, if the difference between the predicted load and the average value of the load at the preset number of time points before calculation is greater than a set difference threshold value, and the predicted load is greater than a set start-stop threshold value, it is a start stage; If the difference between the average value of the load at the preset number of time points before calculation and the average value of the predicted load is greater than a set difference threshold value, or the predicted load is less than or equal to a set start-stop threshold value, it is a stop stage; Otherwise, it is a normal operation stage.

6. The constant temperature and humidity air conditioning system load prediction method based on transfer learning according to claim 5, characterized in that: The predicted load is corrected for shutdown or startup thermal inertia according to the air conditioner start-stop state, specifically: The internal envelope area of the factory and the factory volume are obtained, the internal envelope includes the inner wall and the floor; the corrected load calculation formula is: wherein, is the corrected load; is the predicted load; , are the set-up and shut-down correction factors, respectively; is the time step between two instants; is the thermal inertia constant; is the internal envelope area; is the plant volume.

7. A constant temperature and humidity air conditioning system load prediction system based on the method of any one of claims 1-6, comprising a load prediction model construction module, a benchmark load prediction model acquisition module, a transfer learning module and a correction module, characterized by: The load prediction model construction module: obtains historical factory environmental parameter sequences and corresponding loads under different factory environmental parameters, the historical factory environmental parameter sequences include factory temperature data sequences and factory humidity data sequences within a set period; the feature data of each historical factory environmental parameter sequence is extracted, All feature data is clustered, and the historical factory environmental parameter sequence and its corresponding load in the same cluster are used as a training set to train a load prediction model; The benchmark load prediction model acquisition module: obtains the current factory environmental parameter sequence, calculates the sequence similarity between the current factory environmental parameter sequence and the historical factory environmental parameter sequence corresponding to the cluster center of all clusters, and takes the load prediction model corresponding to the cluster center with the maximum sequence similarity as the benchmark load prediction model; The transfer learning module: takes the benchmark load prediction model as a source domain model, constructs a target domain model with the same network structure as the source domain model, and performs transfer learning on the target domain model; input the current factory environmental parameter sequence into the target domain model to output the predicted load; The correction module: determines the air conditioner start-stop state of the factory at the next time point according to the predicted load, and corrects the predicted load for shutdown or startup thermal inertia according to the air conditioner start-stop state.

Images