A multi-task transfer-based comprehensive energy system load prediction method and system

Abstract

The present application relates to the technical field of load prediction, and provides a comprehensive energy system load prediction method and system based on multi-task transfer, which comprises: acquiring a load data set, updating a basic load prediction model parameter, and calculating an influence coefficient of each model parameter, and transferring the basic load prediction model parameter to a new load prediction model; acquiring load data of a current time in a new comprehensive energy system operation process, calculating a current time cumulative error, and if the cumulative error is greater than a first threshold value, then updating the new load prediction model parameter according to the influence coefficient, and then using the new load prediction model to perform load prediction; putting the load data of the current time into a new comprehensive energy system data set, calculating a distribution difference between the new comprehensive energy system data set and the load data set, and if the distribution difference is greater than a second threshold value, then adding the data in the new comprehensive energy system data set to the load data set, and re-updating the basic load prediction model parameter. The new load prediction model is ensured to learn the knowledge of the basic load prediction model.

Description

Technical Field

[0001] This invention belongs to the field of load forecasting technology, and in particular relates to a method and system for load forecasting of integrated energy systems based on multi-task migration. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Load forecasting, as a key technology supporting the energy dispatch and operation optimization of integrated energy systems, is of great significance for improving the economy and security of integrated energy systems. With the accelerated construction of integrated energy systems, the problems existing in load forecasting of newly constructed energy systems are receiving increasing attention.

[0004] First, there are often significant differences between new and old integrated energy systems in terms of energy consumption patterns, equipment composition, and operation strategies, which leads to differences in the number of load forecasting tasks among different integrated energy systems, thus hindering the effective reuse of existing knowledge and models.

[0005] In addition, due to the combined influence of multiple dynamic factors such as seasonal changes, weather fluctuations, and user energy consumption behavior, the load data distribution of the new integrated energy system exhibits strong time-varying characteristics. This requires the prediction model to not only have good initial fitting ability, but also to be able to optimize and adaptively adjust in real time according to changes in data distribution.

[0006] Currently, online prediction methods targeting the time-varying characteristics of data can continuously update model parameters with the help of real-time data streams to adapt to short-term load fluctuations. However, these methods are often accompanied by the problem of historical knowledge forgetting during the dynamic learning process. That is, while the model adapts to new data, it may gradually lose its memory of past patterns, which leads to a decline in generalization ability and ultimately affects the accuracy and stability of the overall prediction.

[0007] Therefore, maintaining the model's ability to continuously utilize historical information while adapting to the time-varying characteristics of data has become a key challenge in improving the load forecasting performance of newly built integrated energy systems. Summary of the Invention

[0008] To address the technical problems mentioned above, this invention provides a method and system for load forecasting of integrated energy systems based on multi-task migration. First, a basic load forecasting model is trained using data from the source integrated energy system, and the influence coefficient of each parameter in the basic load forecasting model is calculated. When the new load forecasting model of the new integrated energy system is triggered for updating, the regularized loss is calculated based on the parameters of the basic load forecasting model and the influence coefficients, and the parameters of the new load forecasting model are updated. This ensures that the new load forecasting model learns the knowledge of the basic load forecasting model, thereby preventing knowledge forgetting and realizing knowledge transfer across energy systems.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The first aspect of the present invention provides a comprehensive energy system load forecasting method based on multi-task migration, comprising:

[0011] Obtain the load dataset, use the load dataset to update the parameters of the basic load forecasting model, calculate the influence coefficient of each model parameter, and migrate the parameters of the basic load forecasting model to the new load forecasting model.

[0012] Obtain the load data at the current moment during the operation of the new integrated energy system, calculate the cumulative error at the current moment. If the cumulative error is greater than the first threshold, update the parameters of the new load forecasting model according to the influence coefficient, and then use the new load forecasting model to forecast the load. Put the load data at the current moment into the new integrated energy system dataset, calculate the distribution difference between the new integrated energy system dataset and the load dataset. If the distribution difference is greater than the second threshold, add the data in the new integrated energy system dataset to the load dataset, and update the parameters of the basic load forecasting model again.

[0013] Furthermore, both the basic load forecasting model and the new load forecasting model include a feature extraction module and a forecasting module, and the feature extraction module integrates a dependency extraction algorithm that is independent of the number of tasks.

[0014] Furthermore, the influence coefficient of each model parameter is: ;in, This represents the i-th model parameter of the basic load forecasting model. This represents the influence coefficient of the i-th model parameter. This represents the optimal parameters of the basic load forecasting model. Indicates the size of the load dataset. This represents the j-th sample input in the load dataset. This represents the basic load forecasting model.

[0015] Furthermore, the updating of the new load forecasting model parameters based on the influence coefficient is expressed as follows: ;in, These are the true values ​​of the centralized samples in the new integrated energy system dataset. These are the predicted values ​​from the new load forecasting model. It's the learning rate. For the parameters of the new load forecasting model in generation t, For the parameters of the new load forecasting model in generation t+1, B The total number of forecast tasks for the new integrated energy system. L N The loss function for the new load forecasting model is the parameter regularization loss. , This represents the influence coefficient of the i-th model parameter. It is the i-th parameter of the new load forecasting model. It is the i-th optimal parameter of the basic load forecasting model.

[0016] Furthermore, the distribution difference is ; where, function , It's a hyperparameter. This represents the nth sample in the new integrated energy system dataset. This represents the nth sample in the load dataset. This represents the m-th sample in the new integrated energy system dataset. Represents the m-th sample in the load dataset, | D | represents the size of the load dataset, | D N | This represents the size of the new integrated energy system dataset.

[0017] A second aspect of the present invention provides a comprehensive energy system load forecasting system based on multi-task migration, comprising:

[0018] The migration module is configured to: acquire the load dataset, update the parameters of the basic load forecasting model using the load dataset, calculate the influence coefficient of each model parameter, and migrate the parameters of the basic load forecasting model to the new load forecasting model.

[0019] The prediction module is configured to: acquire the load data at the current moment during the operation of the new integrated energy system, calculate the cumulative error at the current moment, and if the cumulative error is greater than the first threshold, update the parameters of the new load prediction model according to the influence coefficient and then use the new load prediction model to predict the load; put the load data at the current moment into the new integrated energy system dataset, calculate the distribution difference between the new integrated energy system dataset and the load dataset, and if the distribution difference is greater than the second threshold, add the data in the new integrated energy system dataset to the load dataset and update the parameters of the basic load prediction model again.

[0020] Furthermore, both the basic load forecasting model and the new load forecasting model include a feature extraction module and a forecasting module, and the feature extraction module integrates a dependency extraction algorithm that is independent of the number of tasks.

[0021] Furthermore, the influence coefficient of each model parameter is: ;in, This represents the i-th model parameter of the basic load forecasting model. This represents the influence coefficient of the i-th model parameter. This represents the optimal parameters of the basic load forecasting model. Indicates the size of the load dataset. This represents the j-th sample input in the load dataset. This represents the basic load forecasting model.

[0022] Furthermore, the updating of the new load forecasting model parameters based on the influence coefficient is expressed as follows: ;in, These are the true values ​​of the centralized samples in the new integrated energy system dataset. These are the predicted values ​​from the new load forecasting model. It's the learning rate. For the parameters of the new load forecasting model in generation t, For the parameters of the new load forecasting model in generation t+1, B The total number of forecast tasks for the new integrated energy system. L N The loss function for the new load forecasting model is the parameter regularization loss. , This represents the influence coefficient of the i-th model parameter. It is the i-th parameter of the new load forecasting model. It is the i-th optimal parameter of the basic load forecasting model.

[0023] Furthermore, the distribution difference is ; where, function , It's a hyperparameter. This represents the nth sample in the new integrated energy system dataset. This represents the nth sample in the load dataset. This represents the m-th sample in the new integrated energy system dataset. Represents the m-th sample in the load dataset, | D | represents the size of the load dataset, | D N | This represents the size of the new integrated energy system dataset.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] This invention first trains a basic load forecasting model using data from a source integrated energy system and calculates the influence coefficient of each parameter in the basic load forecasting model. When the new load forecasting model of the new integrated energy system is triggered for update, the regularized loss is calculated based on the parameters of the basic load forecasting model and the influence coefficient, and the parameters of the new load forecasting model are updated to ensure that the new load forecasting model learns the knowledge of the basic load forecasting model, thereby preventing knowledge forgetting and realizing knowledge transfer across energy systems.

[0026] This invention addresses the issue of different forecasting tasks for various integrated energy systems by designing a load forecasting model to adapt to the load forecasting tasks of different integrated energy systems. Attached Figure Description

[0027] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0028] Figure 1 This is a flowchart of a comprehensive energy system load forecasting method based on multi-task migration, according to Embodiment 1 of the present invention. Detailed Implementation

[0029] 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.

[0030] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0031] Example 1

[0032] This embodiment provides a comprehensive energy system load forecasting method based on multi-task migration.

[0033] This embodiment provides a load forecasting method for integrated energy systems based on multi-task migration. To address the issue of inconsistent load forecasting task numbers across integrated energy systems, a task-adaptive load forecasting model is designed. This load forecasting model includes a feature extraction module and a forecasting module. The feature extraction module integrates a dependency extraction algorithm that is independent of the number of tasks, and can adaptively adjust according to the number of forecasting tasks in different integrated energy systems without changing the structure of the load forecasting model.

[0034] This embodiment provides a multi-task migration-based integrated energy system load forecasting method. To address the knowledge forgetting problem in the online optimization process, a knowledge-transfer-based parameter update strategy is proposed. When a new integrated energy system triggers a model update, this strategy uses parameter regularization loss to transfer the knowledge of the basic load forecasting model to the new load forecasting model, enabling the new load forecasting model to adapt to changes in the load data distribution of the new integrated energy system while preventing knowledge forgetting.

[0035] This embodiment provides a comprehensive energy system load forecasting method based on multi-task migration, such as... Figure 1 As shown, it includes the following steps:

[0036] Step 1: Collect load data from the integrated energy system.

[0037] A comprehensive energy system includes electrical load, cooling load, and heating load. There are several types of loads, and each type of load has... If there are 10 prediction nodes, then the total number of prediction tasks is 10. .

[0038] The load data is segmented into multiple samples. and form a load dataset. D ,in, X It is a sample input. Y It is the sample output. L Indicates the length of historical load data.

[0039] Step 2: Normalize the load data for subsequent updates to the basic load forecasting model (hereinafter referred to as the basic model).

[0040] Step 3: Construction of the basic load forecasting model.

[0041] Basic load forecasting model Divided into feature extraction modules and prediction module .

[0042] (1) Feature extraction module The features used to extract input load data are calculated as follows:

[0043] (1);

[0044] in, and Represents a nonlinear function. Represents the normalization function. H 1. H 2 and H Indicates intermediate features. Represents the dot product. For learnable parameters, d Representing feature dimension, h Represents the characteristic matrix, h i express h The i-th eigenvector in The calculation method is as follows:

[0045] (2);

[0046] in, Represents the load input matrix The i-th and j-th load vectors in the middle, ; , All are learnable parameters; Represents an exponential function; This indicates the degree of dependency between the i-th and j-th loads; This is valid only when load i and load j are of the same type; otherwise, it is 0. This is valid only if load i and load j are different types of loads; otherwise, it is 0.

[0047] (2) Prediction module Used to output predicted values, its calculation method is as follows:

[0048] (3);

[0049] in, These are the predicted values ​​and the learnable parameters.

[0050] It is worth noting that the calculation of the basic load forecasting model does not depend on the number of tasks. A This means that the basic load forecasting model can adapt to any number of load forecasting tasks.

[0051] Step 4: Update the parameters of the basic load forecasting model and calculate the impact coefficient.

[0052] Using load datasets D Data updates the parameters of the basic load forecasting model. :

[0053] (4);

[0054] in, Y It is a dataset D The true value in These are the predicted values ​​from the basic load forecasting model. It's the learning rate. The parameters of the basic load forecasting model for generation t are... These are the parameters of the basic load forecasting model for generation t+1. This is the loss function.

[0055] Then, calculate the influence coefficient of each parameter:

[0056] (5);

[0057] in, This represents the i-th parameter of the basic load forecasting model. This represents the influence coefficient of the i-th parameter. This represents the optimal parameters of the basic load forecasting model. Indicates the size of the dataset. Represents the dataset D The j-th sample input, This represents the basic load forecasting model.

[0058] Step 5: Transfer the parameters of the basic load forecasting model to the new load forecasting model.

[0059] The new integrated energy system (hereinafter referred to as the new system) includes There are several types of loads, and each type of load has... If there are prediction nodes, then the total number of prediction tasks is .

[0060] New load forecasting models in new integrated energy systems Includes feature extraction module and prediction module .

[0061] The new load forecasting model after migration can directly perform tasks in the new integrated energy system.

[0062] Step 6: Load data at time t is collected during the operation of the new integrated energy system. .

[0063] Step 7: If there is a predicted value for the new integrated energy system at time t-1... Then, the cumulative error at time t is calculated as follows:

[0064] (6);

[0065] in, E t-1 This represents the cumulative error at time t-1.

[0066] Step 8: If the accumulated error E t Greater than the first threshold th eThen the parameters of the new load forecasting model (i.e., the model of the new system) are optimized.

[0067] Among them, the first threshold th e The settings are based on the mean absolute percentage error (MAPE), which is the ratio of the error to the true value. Generally, it is considered that the predicted data is usable when the MAPE is less than 10%.

[0068] From the new integrated energy system dataset K samples are extracted from the data, and then the K extracted samples are input into the new load prediction model. In the middle, the predicted output is obtained as Based on the influence coefficient, the loss of parameter regularization is calculated. :

[0069] (7);

[0070] in, It is the i-th parameter of the new load forecasting model. It is the i-th optimal parameter of the basic load forecasting model.

[0071] L M Regular constraints are used to allow the new load forecasting model to learn from the basic load forecasting model, thereby preventing the forgetting of historical knowledge.

[0072] Then, the parameters of the new load forecasting model are optimized using the calculation method in the following formula. :

[0073] (8);

[0074] in, These are the true values ​​of the centralized samples in the new integrated energy system dataset. These are the predicted values ​​from the new load forecasting model. It's the learning rate. For the parameters of the new load forecasting model in generation t, For the parameters of the new load forecasting model in generation t+1, L N This is the loss function for the new load forecasting model.

[0075] Step 9: Construct the input data at time t and predict the future load value.

[0076] Using load data The input matrix is ​​formed by combining the load data from time t-L+1 to time t-1. Input the input matrix into the new load forecasting model. In the middle, the output is obtained As the final load forecast output ,in, This represents the load forecast for time t+1 at time t.

[0077] Step 10: Load data Add to the new integrated energy system dataset Then calculate and Distribution differences between M :

[0078] (9);

[0079] Among them, the function , It's a hyperparameter. Represents the new integrated energy system dataset The nth sample in Represents the load dataset The nth sample in Represents the new integrated energy system dataset The m-th sample in the series, Represents the load dataset The m-th sample in the dataset.

[0080] If M is greater than the second threshold th m Then, the new integrated energy system dataset The data in the dataset is merged into the load dataset. Then return to step 4 to re-optimize the basic load forecasting model.

[0081] Among them, the second threshold th m Influenced by factors such as hyperparameter selection and data dimensionality.

[0082] This embodiment provides a load forecasting method for integrated energy systems based on multi-task migration. The load forecasting model consists of a feature extraction module and a forecasting module, which are used for load feature extraction and predicting future load values, respectively, and can adapt to the load forecasting tasks of different integrated energy systems. In addition, the feature extraction module is built based on a specially designed dependency extraction method, which can extract the relationship between different loads based on load data.

[0083] This embodiment provides a load forecasting method for integrated energy systems based on multi-task migration. First, a basic load forecasting model is trained using data from the source integrated energy system, and the influence coefficient of each parameter in the basic load forecasting model is calculated. When the new load forecasting model of the new integrated energy system is triggered for update, the regularized loss is calculated based on the parameters of the basic load forecasting model and the influence coefficient, and the parameters of the new load forecasting model are updated to ensure that the new load forecasting model learns the knowledge of the basic load forecasting model, thereby preventing the forgetting of knowledge.

[0084] This embodiment provides a comprehensive energy system load forecasting method based on multi-task migration, which can realize knowledge transfer across energy systems and can be used for load forecasting tasks of comprehensive energy systems.

[0085] This embodiment provides a load forecasting method for integrated energy systems based on multi-task migration. To address the issue of different forecasting tasks for different integrated energy systems, a load forecasting model is designed to adapt to the load forecasting tasks of different integrated energy systems.

[0086] This embodiment provides a comprehensive energy system load forecasting method based on multi-task migration. It designs an optimization strategy based on knowledge transfer, which supports the model to be dynamically updated according to real-time data streams, while ensuring that the knowledge learned from the original load data is retained.

[0087] Example 2

[0088] This embodiment provides a comprehensive energy system load forecasting system based on multi-task migration, comprising:

[0089] The migration module is configured to: acquire the load dataset, update the parameters of the basic load forecasting model using the load dataset, calculate the influence coefficient of each model parameter, and migrate the parameters of the basic load forecasting model to the new load forecasting model.

[0090] The prediction module is configured to: acquire the load data at the current moment during the operation of the new integrated energy system, calculate the cumulative error at the current moment, and if the cumulative error is greater than the first threshold, update the parameters of the new load prediction model according to the influence coefficient and then use the new load prediction model to predict the load; put the load data at the current moment into the new integrated energy system dataset, calculate the distribution difference between the new integrated energy system dataset and the load dataset, and if the distribution difference is greater than the second threshold, add the data in the new integrated energy system dataset to the load dataset and update the parameters of the basic load prediction model again.

[0091] Furthermore, both the basic load forecasting model and the new load forecasting model include a feature extraction module and a forecasting module, and the feature extraction module integrates a dependency extraction algorithm that is independent of the number of tasks.

[0092] Furthermore, the influence coefficient of each model parameter is: ;in, This represents the i-th model parameter of the basic load forecasting model. This represents the influence coefficient of the i-th model parameter. This represents the optimal parameters of the basic load forecasting model. Indicates the size of the load dataset. This represents the j-th sample input in the load dataset. This represents the basic load forecasting model.

[0093] Furthermore, the updating of the new load forecasting model parameters based on the influence coefficient is expressed as follows: ;in, These are the true values ​​of the centralized samples in the new integrated energy system dataset. These are the predicted values ​​from the new load forecasting model. It's the learning rate. For the parameters of the new load forecasting model in generation t, For the parameters of the new load forecasting model in generation t+1, B The total number of forecast tasks for the new integrated energy system. L N The loss function for the new load forecasting model is the parameter regularization loss. , This represents the influence coefficient of the i-th model parameter. It is the i-th parameter of the new load forecasting model. It is the i-th optimal parameter of the basic load forecasting model.

[0094] Furthermore, the distribution difference is ; where, function , It's a hyperparameter. This represents the nth sample in the new integrated energy system dataset. This represents the nth sample in the load dataset. This represents the m-th sample in the new integrated energy system dataset. Represents the m-th sample in the load dataset, | D | represents the size of the load dataset, | D N | This represents the size of the new integrated energy system dataset.

[0095] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A comprehensive energy system load forecasting method based on multi-task migration, characterized in that, include: Construction of the basic load forecasting model: Basic load forecasting model Divided into feature extraction modules and prediction module ; (1) Feature extraction module The features used to extract input load data are calculated as follows: (1)； in, and Represents a nonlinear function. Represents the normalization function. H 1. H 2 and H Indicates intermediate features. Represents the dot product. For learnable parameters, d Representing feature dimension, h Represents the characteristic matrix, h i express h The i-th eigenvector in The calculation method is as follows: (2)； in, Represents the load input matrix The i-th and j-th load vectors in the middle, ; , All are learnable parameters; Represents an exponential function; This indicates the degree of dependency between the i-th and j-th loads; This is valid only when load i and load j are of the same type; otherwise, it is 0. This is valid only when load i and load j are different types of loads; otherwise, it is 0. (2) Prediction module Used to output predicted values, its calculation method is as follows: (3)； in, For predicted values, learnable parameters; It is worth noting that the calculation of the basic load forecasting model does not depend on the number of tasks. A That is, the basic load forecasting model can adapt to any number of load forecasting tasks; Obtain the load dataset, use the load dataset to update the parameters of the basic load forecasting model, calculate the influence coefficient of each model parameter, and migrate the parameters of the basic load forecasting model to the new load forecasting model. The influence coefficient of each model parameter is: ;in, This represents the i-th model parameter of the basic load forecasting model. This represents the influence coefficient of the i-th model parameter. This represents the optimal parameters of the basic load forecasting model. Indicates the size of the load dataset. This represents the j-th sample input in the load dataset. This represents the basic load forecasting model; Obtain the load data at the current moment during the operation of the new integrated energy system, calculate the cumulative error at the current moment, and if the cumulative error is greater than the first threshold, update the parameters of the new load prediction model according to the influence coefficient, and then use the new load prediction model to predict the load. The update of the new load forecasting model parameters based on the influence coefficient is expressed as follows: ;in, These are the true values ​​of the centralized samples in the new integrated energy system dataset. These are the predicted values ​​from the new load forecasting model. It's the learning rate. For the parameters of the new load forecasting model in generation t, For the parameters of the new load forecasting model in generation t+1, B The total number of forecast tasks for the new integrated energy system. L N The loss function for the new load forecasting model is the parameter regularization loss. , This represents the influence coefficient of the i-th model parameter. It is the i-th parameter of the new load forecasting model. The i-th optimal parameter of the basic load forecasting model; The load data at the current moment is put into the new integrated energy system dataset. The distribution difference between the new integrated energy system dataset and the load dataset is calculated. If the distribution difference is greater than the second threshold, the data in the new integrated energy system dataset is added to the load dataset, and the parameters of the basic load forecasting model are updated again.

2. The integrated energy system load forecasting method based on multi-task migration as described in claim 1, characterized in that, Both the basic load forecasting model and the new load forecasting model include a feature extraction module and a forecasting module, and the feature extraction module integrates a dependency extraction algorithm that is independent of the number of tasks.

3. The integrated energy system load forecasting method based on multi-task migration as described in claim 1, characterized in that, The distribution difference is ; where, function , It's a hyperparameter. This represents the nth sample in the new integrated energy system dataset. This represents the nth sample in the load dataset. This represents the m-th sample in the new integrated energy system dataset. Represents the m-th sample in the load dataset, | D | represents the size of the load dataset, | D N | This represents the size of the new integrated energy system dataset.

4. A comprehensive energy system load forecasting system based on multi-task migration, characterized in that, include: Construction of the basic load forecasting model: Basic load forecasting model Divided into feature extraction modules and prediction module ; (1) Feature extraction module The features used to extract input load data are calculated as follows: (1)； in, and Represents a nonlinear function. Represents the normalization function. H 1. H 2 and H Indicates intermediate features. Represents the dot product. For learnable parameters, d Representing feature dimension, h Represents the characteristic matrix, h i express h The i-th eigenvector in The calculation method is as follows: (2)； in, Represents the load input matrix The i-th and j-th load vectors in the middle, ; , All are learnable parameters; Represents an exponential function; This indicates the degree of dependency between the i-th and j-th loads; This is valid only when load i and load j are of the same type; otherwise, it is 0. This is valid only when load i and load j are different types of loads; otherwise, it is 0. (2) Prediction module Used to output predicted values, its calculation method is as follows: (3)； in, For predicted values, learnable parameters; It is worth noting that the calculation of the basic load forecasting model does not depend on the number of tasks. A That is, the basic load forecasting model can adapt to any number of load forecasting tasks; The migration module is configured to: acquire the load dataset, update the parameters of the basic load forecasting model using the load dataset, calculate the influence coefficient of each model parameter, and migrate the parameters of the basic load forecasting model to the new load forecasting model. The influence coefficient of each model parameter is: ;in, This represents the i-th model parameter of the basic load forecasting model. This represents the influence coefficient of the i-th model parameter. This represents the optimal parameters of the basic load forecasting model. Indicates the size of the load dataset. This represents the j-th sample input in the load dataset. This represents the basic load forecasting model; The prediction module is configured to: acquire the load data at the current moment during the operation of the new integrated energy system, calculate the cumulative error at the current moment, and if the cumulative error is greater than the first threshold, update the parameters of the new load prediction model according to the influence coefficient and then use the new load prediction model to predict the load. The update of the new load forecasting model parameters based on the influence coefficient is expressed as follows: ;in, These are the true values ​​of the centralized samples in the new integrated energy system dataset. These are the predicted values ​​from the new load forecasting model. It's the learning rate. For the parameters of the new load forecasting model in generation t, For the parameters of the new load forecasting model in generation t+1, B The total number of forecast tasks for the new integrated energy system. L N The loss function for the new load forecasting model is the parameter regularization loss. , This represents the influence coefficient of the i-th model parameter. It is the i-th parameter of the new load forecasting model. The i-th optimal parameter of the basic load forecasting model; The load data at the current moment is put into the new integrated energy system dataset. The distribution difference between the new integrated energy system dataset and the load dataset is calculated. If the distribution difference is greater than the second threshold, the data in the new integrated energy system dataset is added to the load dataset, and the parameters of the basic load forecasting model are updated again.

5. The integrated energy system load forecasting system based on multi-task migration as described in claim 4, characterized in that, Both the basic load forecasting model and the new load forecasting model include a feature extraction module and a forecasting module, and the feature extraction module integrates a dependency extraction algorithm that is independent of the number of tasks.

6. The integrated energy system load forecasting system based on multi-task migration as described in claim 4, characterized in that, The influence coefficient of each model parameter is: ;in, This represents the i-th model parameter of the basic load forecasting model. This represents the influence coefficient of the i-th model parameter. This represents the optimal parameters of the basic load forecasting model. Indicates the size of the load dataset. This represents the j-th sample input in the load dataset. This represents the basic load forecasting model.

7. The integrated energy system load forecasting system based on multi-task migration as described in claim 4, characterized in that, The update of the new load forecasting model parameters based on the influence coefficient is expressed as follows: ;in, These are the true values ​​of the centralized samples in the new integrated energy system dataset. These are the predicted values ​​from the new load forecasting model. It's the learning rate. For the parameters of the new load forecasting model in generation t, For the parameters of the new load forecasting model in generation t+1, B The total number of forecast tasks for the new integrated energy system. L N The loss function for the new load forecasting model is the parameter regularization loss. , This represents the influence coefficient of the i-th model parameter. It is the i-th parameter of the new load forecasting model. It is the i-th optimal parameter of the basic load forecasting model.

8. The integrated energy system load forecasting system based on multi-task migration as described in claim 4, characterized in that, The distribution difference is ; where, function , It's a hyperparameter. This represents the nth sample in the new integrated energy system dataset. This represents the nth sample in the load dataset. This represents the m-th sample in the new integrated energy system dataset. Represents the m-th sample in the load dataset, | D | represents the size of the load dataset, | D N | This represents the size of the new integrated energy system dataset.

Images