Load prediction method, load prediction model training method, and related device

Abstract

The disclosure provides a load prediction method, a load prediction model training method and related equipment, which relate to the field of artificial intelligence such as deep learning, and the method comprises the following steps: obtaining historical load time series data, historical exogenous time series data and predictable exogenous time series data in a prediction time range; obtaining first load change information of the historical load time series data under the influence of the historical exogenous time series data, and extracting first features corresponding to the first load change information; obtaining first predicted load time series data in the prediction time range under the influence of the predictable exogenous time series data according to the first features; and obtaining target predicted load time series data in the prediction time range according to the first predicted load time series data. In the disclosure, the load prediction based on the exogenous time series data is realized, the accuracy of the load prediction is improved, the load prediction method has better applicability and practicability, and the load prediction method is optimized.

Description

Technical Field

[0001] This disclosure relates to the field of data processing, particularly to the field of artificial intelligence such as deep learning, and is applicable to power load forecasting scenarios. Background Technology

[0002] With the development of technology, power load forecasting plays an increasingly important role in power dispatching and operation. In practice, accurate power load forecasting can provide important data support for power unit start-up and shutdown, economic dispatching, and load management.

[0003] In practice, many factors influence load data fluctuations, and these factors are complex. Existing technologies can use model algorithms to process and analyze historical load time-series data for load forecasting, but this ignores other factors affecting load fluctuations, resulting in lower accuracy of load forecasts. Summary of the Invention

[0004] This disclosure proposes a load forecasting method, a training method for a load forecasting model, and related equipment.

[0005] According to a first aspect of this disclosure, a load forecasting method is proposed, the method comprising: acquiring historical load time-series data, historical exogenous time-series data, and predictable exogenous time-series data within a forecast time range; acquiring first load change information of the historical load time-series data under the influence of the historical exogenous time-series data, and extracting a first feature corresponding to the first load change information; acquiring first predicted load time-series data within the forecast time range under the influence of the predictable exogenous time-series data based on the first feature; and acquiring target predicted load time-series data within the forecast time range based on the first predicted load time-series data.

[0006] According to a second aspect of this disclosure, a training method for a load forecasting model is proposed. The method includes: acquiring sample load time-series data and sample exogenous time-series data within a sample time range; inputting the sample load time-series data and sample exogenous time-series data into a load forecasting model to be trained, wherein the load forecasting model acquires first sample load change information of the sample load time-series data under the influence of the sample exogenous time-series data, and extracts first sample features of the first sample load change information; based on the first sample features, acquiring first load time-series data within the sample time range under the influence of the sample exogenous time-series data; adjusting the load forecasting model according to the first load time-series data and the sample load time-series data, and returning to continue training the adjusted load forecasting model to obtain a trained target load forecasting model.

[0007] According to a third aspect of this disclosure, a load time-series forecasting method is proposed, the method comprising: acquiring historical load time-series data to be predicted, and predictable exogenous time-series data within a prediction time range; inputting the historical load time-series data and the predictable exogenous time-series data into a trained target load forecasting model; and obtaining, from the output of the target load forecasting model, target predicted load time-series data within the prediction time range, based on the historical load time-series data and the predictable exogenous time-series data; wherein the target load forecasting model is obtained based on the training method of the load forecasting model described in the second aspect above.

[0008] According to a fourth aspect of this disclosure, a load forecasting device is proposed, the device comprising: a first acquisition module, configured to acquire historical load time-series data, historical exogenous time-series data, and predictable exogenous time-series data within a forecast time range; a first extraction module, configured to acquire first load change information of the historical load time-series data under the influence of the historical exogenous time-series data, and extract a first feature corresponding to the first load change information; a first forecasting module, configured to acquire first forecast load time-series data within the forecast time range under the influence of the predictable exogenous time-series data based on the first feature; and a second forecasting module, configured to acquire target forecast load time-series data within the forecast time range based on the first forecast load time-series data.

[0009] According to a fifth aspect of this disclosure, a training apparatus for a load forecasting model is proposed. The apparatus includes: a second acquisition module for acquiring sample load time-series data and sample exogenous time-series data within a sample time range; a second extraction module for inputting the sample load time-series data and sample exogenous time-series data into a load forecasting model to be trained, wherein the load forecasting model acquires first sample load change information of the sample load time-series data under the influence of the sample exogenous time-series data, and extracts a first sample feature of the first sample load change information; a third prediction module for acquiring first load time-series data within the sample time range under the influence of the sample exogenous time-series data based on the first sample feature; and a training module 94 for adjusting the load forecasting model according to the first load time-series data and the sample load time-series data, and returning to continue training the adjusted load forecasting model to obtain a trained target load forecasting model.

[0010] According to a sixth aspect of this disclosure, a load time series forecasting device is proposed, the device comprising: a third acquisition module, configured to acquire historical load time series data to be predicted, and predictable exogenous time series data within a prediction time range; a fourth prediction module, configured to input the historical load time series data and the predictable exogenous time series data into a trained target load forecasting model, and acquire, from the output of the target load forecasting model, target predicted load time series data within the prediction time range, based on the historical load time series data and the predictable exogenous time series data; wherein the target load forecasting model is obtained based on the training device for the load forecasting model described in the fifth aspect above.

[0011] According to a seventh aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the load forecasting method of the first aspect, the load forecasting model training method of the second aspect, and the load time-series forecasting method of the third aspect.

[0012] According to the eighth aspect of this disclosure, a non-transient computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause the computer to perform the load forecasting method described in the first aspect, the load forecasting model training method described in the second aspect, and the load timing forecasting method described in the third aspect.

[0013] According to the ninth aspect of this disclosure, a computer program product is proposed, comprising a computer program that, when executed by a processor, implements the load forecasting method described in the first aspect, the load forecasting model training method described in the second aspect, and the load time-series forecasting method described in the third aspect.

[0014] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0015] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0016] Figure 1 This is a schematic flowchart of a load forecasting method according to an embodiment of the present disclosure;

[0017] Figure 2 This is a schematic flowchart of a load forecasting method according to another embodiment of the present disclosure;

[0018] Figure 3 This is a flowchart illustrating a training method for a load prediction model according to another embodiment of the present disclosure.

[0019] Figure 4 This is a flowchart illustrating a training method for a load prediction model according to another embodiment of the present disclosure.

[0020] Figure 5 This is a flowchart illustrating a training method for a load prediction model according to another embodiment of the present disclosure.

[0021] Figure 6 This is a flowchart illustrating a training method for a load prediction model according to another embodiment of the present disclosure.

[0022] Figure 7 This is a flowchart illustrating a training method for a load prediction model according to an embodiment of the present disclosure.

[0023] Figure 8 This is a schematic diagram of a load forecasting model according to an embodiment of the present disclosure;

[0024] Figure 9 This is a schematic flowchart of a load timing prediction method according to an embodiment of the present disclosure;

[0025] Figure 10 This is a schematic diagram illustrating the model performance of a target load time-series prediction model according to an embodiment of the present disclosure;

[0026] Figure 11 This is a schematic diagram of the structure of a load forecasting device according to an embodiment of the present disclosure;

[0027] Figure 12 This is a schematic diagram of the structure of a training device for a load prediction model according to an embodiment of the present disclosure;

[0028] Figure 13 This is a schematic diagram of the structure of a load timing prediction device according to an embodiment of the present disclosure;

[0029] Figure 14 This is a schematic block diagram of an electronic device according to an embodiment of the present disclosure. Detailed Implementation

[0030] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0031] Deep learning (DL) is a new research direction in the field of machine learning (ML). It was introduced into machine learning to bring it closer to its original goal—artificial intelligence. Deep learning learns the inherent laws and hierarchical representations of sample data. The information gained during this learning process greatly aids in the interpretation of data such as text, images, and sound. Its ultimate goal is to enable machines to possess analytical and learning capabilities like humans, capable of recognizing data such as text, images, and sound. Deep learning is a complex machine learning algorithm that has achieved results in speech and image recognition far exceeding previous related technologies.

[0032] Artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It involves both hardware and software technologies. AI hardware technologies generally include computer vision, speech recognition, natural language processing, and related technologies such as deep learning, big data processing, and knowledge graphs.

[0033] Figure 1 This is a schematic flowchart of a load forecasting method according to an embodiment of the present disclosure, as shown below. Figure 1 As shown, the method includes:

[0034] S101, acquire historical load time series data, historical exogenous time series data, and predictable exogenous time series data within the prediction time range.

[0035] In practice, changes in electricity load may be affected by external factors. Therefore, in the process of forecasting electricity load, we can combine historical load data within a historical time range with data on external factors that affect changes in historical load data within the historical time range to achieve load forecasting for the forecast time range.

[0036] For example, if the weather is different on dates 1 and 2 in the same season, the electricity load data on date 1 may differ from that on date 2. In this scenario, it can be determined that the weather has a certain degree of influence on the change in electricity load.

[0037] Specifically, load time series data within a set historical time range can be read from the set location of the power load data storage and determined as the historical load time series data required for load forecasting within the forecast time range.

[0038] Correspondingly, time-series data corresponding to exogenous factors within the same historical time range can also be obtained from relevant storage locations as historical exogenous time-series data required for load forecasting within the forecast time range.

[0039] In this embodiment of the disclosure, in order to achieve accurate load forecasting within the forecast time range, predictable time-series data of exogenous factors that may affect load changes in electricity can also be obtained for the forecast time range as predictable exogenous time-series data required for load forecasting.

[0040] Optionally, the foreseeable exogenous time series data may include multiple types of exogenous data. The corresponding type of exogenous time series data can be obtained from the monitoring system to which each type of exogenous data belongs, and the multiple types of exogenous time series data obtained can be integrated to obtain the foreseeable exogenous time series data required for load forecasting.

[0041] S102, obtain the first load change information of historical load time series data under the influence of historical exogenous time series data, and extract the first feature corresponding to the first load change information.

[0042] In this embodiment of the disclosure, the changes in historical load in the historical load time series data under the influence of historical exogenous time series data can be obtained from historical exogenous time series data and historical load time series data.

[0043] Optionally, data analysis can be performed on historical exogenous time series data and historical load time series data, and the results of the data analysis can be used to obtain information on the changes in historical load data under the influence of historical exogenous time series data.

[0044] Among them, the obtained change information is identified as the first load change information of historical load time series data under the influence of historical exogenous time series data.

[0045] It should be noted that data analysis algorithms in related technologies can be used to analyze historical exogenous time-series data and historical load time-series data to obtain the corresponding first load change information; no specific limitations are made here.

[0046] Furthermore, feature extraction is performed on the first load change information. Based on the result of feature extraction, the change characteristics corresponding to the load changes caused by the historical load data in the historical load time series data under the influence of historical exogenous time series data are determined, and these characteristics are identified as the first feature corresponding to the first load change information.

[0047] It should be noted that the first load change information can be extracted based on the feature extraction algorithm in related technologies to obtain the corresponding first feature, without any specific limitations here.

[0048] S103, Based on the first feature, obtain the first predicted load time series data within the predicted time range under the influence of predictable exogenous time series data.

[0049] In this embodiment of the disclosure, the first feature can describe the change characteristics of historical load data in historical load time series data under the influence of historical exogenous time series data.

[0050] This can be understood as, based on the first feature, we can obtain the relevant characteristic patterns of changes in historical load data in historical load time series data under the influence of historical exogenous time series data.

[0051] Therefore, based on the first feature, predictable exogenous data can be analyzed and predicted to obtain the predicted load data that may be reached under the influence of predictable exogenous time series data within the prediction time range, which can be used as the first predicted load time series data within the prediction time range.

[0052] Optionally, a first timestamp included in the predictable exogenous time series data can be obtained, and a corresponding second timestamp can be determined within the prediction time range. Based on the first feature, the predicted load data at the corresponding second timestamp under the influence of the exogenous data at the first timestamp can be obtained, thereby obtaining the first predicted load time series data within the prediction time range.

[0053] S104, Based on the first predicted load time series data, obtain the target predicted load time series data within the predicted time range.

[0054] In this embodiment of the disclosure, the first predicted load time series data can be understood as the load time series data that the power may reach within the predicted time range under the influence of external factors.

[0055] In some implementations, in addition to external factors affecting changes in electricity load, other factors may also influence changes in electricity load.

[0056] In this scenario, the predicted load data that the power may reach under the influence of other factors within the prediction time range can be obtained and integrated with the first predicted load time series data. Based on the integrated time series data, the final target predicted load time series data within the prediction time range can be obtained.

[0057] The load forecasting method proposed in this disclosure acquires historical load time-series data and historical exogenous time-series data, and acquires predictable exogenous time-series data within the forecast time range. It acquires first load change information of historical load data from the historical load time-series data under the influence of historical exogenous time-series data, and obtains a corresponding first feature based on the first load change information. Further, based on the first feature, it acquires first predicted load time-series data within the forecast time range under the influence of predictable exogenous time-series data, and obtains target predicted load time-series data within the forecast time range based on the first predicted load time-series data. In this disclosure, load forecasting within the forecast time range is performed based on the first feature, realizing load forecasting based on exogenous time-series data, improving the accuracy of load forecasting, making the load forecasting method more applicable and practical, and optimizing the load forecasting method.

[0058] In the above embodiments, the acquisition of target predicted load time series data can be combined with... Figure 2 As shown in the next figure, Figure 2 This is a schematic flowchart of a load forecasting method according to another embodiment of the present disclosure, as shown below. Figure 2 As shown, the method includes:

[0059] S201, Obtain the historical time-series characteristics corresponding to the historical load time-series data, wherein the historical time-series characteristics include at least one of historical time-series trend characteristics and historical seasonal characteristics.

[0060] In this embodiment of the disclosure, the variation characteristics of historical load data with time changes within the corresponding historical time range can also be obtained from historical load time series data.

[0061] Optionally, features can be extracted from historical load time series data in the time series trend dimension to obtain the features of historical load time series data in the time series trend dimension, and these features can be identified as the historical time series trend features corresponding to the historical load time series data. Correspondingly, features can also be extracted from historical load time series data in the seasonality dimension to obtain the features of historical load time series data in the seasonality dimension, and these features can be identified as the historical seasonal features of historical load time series data.

[0062] Furthermore, the historical time-series trend characteristics and historical seasonal characteristics of the acquired historical load time-series data are determined as the historical time-series features of the historical load time-series data.

[0063] It should be noted that historical time series features can be historical time series trend features, historical seasonal features, or both historical time series trend features and historical seasonal features. No specific limitation is made here.

[0064] S202, based on historical time series characteristics, obtain the second predicted load time series data within the predicted time range.

[0065] One possible approach is to obtain the predicted load time series data corresponding to the time series trend dimension within the predicted time range based on the historical time series characteristics.

[0066] Accordingly, based on the historical seasonal trend characteristics in the historical time series features, the predicted load time series data corresponding to the seasonal dimension within the prediction time range are obtained.

[0067] Furthermore, based on the predicted load time series data corresponding to the obtained time series trend dimension and the predicted load time series data corresponding to the seasonal dimension, a second predicted load time series data within the predicted time range based on historical time series characteristics is obtained.

[0068] As another possible approach, load forecasting for the forecast time range can be achieved by combining predictable exogenous time-series data with historical time-series data.

[0069] Optionally, historical exogenous features corresponding to historical exogenous time series data can be obtained.

[0070] In this embodiment of the disclosure, feature extraction can be performed on historical exogenous time series data. Specifically, the historical exogenous time series data can be processed by feature extraction algorithms in related technologies, and the exogenous data features corresponding to the historical exogenous time series data can be obtained based on the results of the algorithm processing.

[0071] Among them, the exogenous data features corresponding to the acquired historical exogenous time series data can be determined as the historical exogenous features corresponding to the historical exogenous time series data.

[0072] Furthermore, the second load change information of historical load time series data under the combined influence of historical time series characteristics and historical exogenous characteristics is obtained, and the second feature corresponding to the second load change information is extracted.

[0073] Optionally, load change information under the combined influence of historical time-series characteristics and historical exogenous characteristics in historical time-series features can be obtained from historical load time-series data and determined as the second load change information.

[0074] Correspondingly, it is also possible to obtain load change information under the combined influence of historical seasonality and historical exogenous characteristics in historical load time series data, and to identify it as the second load change information.

[0075] Furthermore, the second load change information is extracted using a feature extraction algorithm in related technologies, thereby obtaining the second feature corresponding to the second load change information.

[0076] It should be noted that the corresponding second load change information can be obtained based on historical time-series trend characteristics and historical exogenous characteristics, or based on historical seasonal characteristics and historical exogenous characteristics, or simultaneously based on historical time-series trend characteristics, historical seasonal characteristics, and historical exogenous characteristics. No specific limitations are made here.

[0077] Furthermore, based on the second feature, second predicted load time series data within the predicted time range are obtained.

[0078] In this embodiment of the disclosure, the second feature can be understood as the change characteristics of historical load data in historical load time series data under the combined influence of historical time series characteristics and historical exogenous characteristics.

[0079] Optionally, the second feature can be used to describe the changes in the predicted load data within the prediction time range under the combined influence of temporal and exogenous features.

[0080] Specifically, the predicted time series features corresponding to the predicted time range can be obtained, wherein the predicted time series features include at least one of the predicted time series trend features and the predicted seasonality features corresponding to the predicted time range.

[0081] In this embodiment of the disclosure, time series data corresponding to the prediction time range can be obtained and determined as prediction time series data, and the prediction time series data corresponding to the prediction time range can be processed by the algorithm based on the feature extraction algorithm of time series data in related technologies.

[0082] Furthermore, based on the results of the algorithm processing, the time series data features corresponding to the predicted time series data corresponding to the predicted time range are obtained and determined as the predicted time series features.

[0083] Among them, the predicted time series data corresponding to the predicted time range can be processed by the algorithm for extracting time series trend features based on the relevant technology, so as to obtain the predicted time series trend features corresponding to the predicted time range.

[0084] Correspondingly, seasonal feature extraction algorithms from related technologies can be used to process the predicted time series data corresponding to the predicted time range, thereby obtaining the predicted seasonal features corresponding to the predicted time range.

[0085] It should be noted that the predicted time series features can be predicted time series trend features, predicted seasonal features, or both predicted time series trend features and predicted seasonal features. No specific limitation is made here.

[0086] Furthermore, predictable exogenous characteristics within the predicted time range are obtained from predictable exogenous time series data.

[0087] In this embodiment of the disclosure, different types of exogenous time series data can be acquired to obtain the predictable exogenous time series data required for load forecasting.

[0088] It can acquire time series data of observed variables, time series data of predicted covariates, and time series data of static variables within the prediction time range.

[0089] Optionally, the time series data of the observed variables may include voltage time series data of electricity, reactive power time series data of electricity, and measured weather time series data within the prediction time range.

[0090] Optionally, the time series data for predicting covariates may include predicted weather time series data within the prediction time range and predicted time series data within the prediction time range.

[0091] Optionally, the static variable time series data may include time series data corresponding to the installed capacity of the motor and time series data corresponding to the bus identification document (ID).

[0092] Furthermore, based on the time series information within the prediction time range, the time series data of observed variables, the predicted exogenous time series data, and the time series data of static variables are integrated along the time series dimension to obtain predictable exogenous time series data within the prediction time range.

[0093] Optionally, observed variable data, predicted exogenous data, and static variable data belonging to the same time series information within the prediction time range can be integrated to obtain predictable exogenous data on the same time series information.

[0094] Furthermore, all predictable exogenous data in the time series information are integrated along the time series dimension, and predictable exogenous time series data for the prediction time range are obtained based on the integrated results.

[0095] Furthermore, based on the second feature, second predicted load time series data within the predicted time range are obtained under the combined influence of predicted time series features and predictable exogenous features.

[0096] Optionally, the second feature can be used to describe the change characteristics of the predicted load data within the prediction time range, under the combined influence of the prediction time series features and the predictable exogenous features.

[0097] Therefore, load forecasting can be performed based on the second feature, according to the forecast time series features and the predictable exogenous features, and the forecasted load data obtained can be determined as the second load forecast data within the forecast time range.

[0098] S203, based on the first predicted load time series data and the second predicted load time series data, obtain the target predicted load time series data.

[0099] In this embodiment of the disclosure, load forecasting can be performed within the forecast time range by combining exogenous influencing factors and temporal influencing factors. Specifically, the first predicted load time series data and the second predicted load time series data can be integrated according to the time series to obtain integrated load time series data, which is used as the target predicted load time series data.

[0100] Optionally, time-series information within the forecast time range can be obtained, and the corresponding first forecast load data and second forecast load data on the same time-series information can be integrated to obtain the target forecast load data on the same time-series information.

[0101] Furthermore, the target predicted load data corresponding to all time series information within the prediction time range are integrated based on the time series within the prediction time range to obtain the final predicted load data predicted based on the predictable exogenous characteristics of predictable exogenous data and the predicted time series characteristics of the prediction time range within the prediction time range, and this data is determined as the target predicted load time series data within the prediction time range.

[0102] The load forecasting method proposed in this disclosure acquires historical time-series characteristics of historical load time-series data and historical exogenous characteristics of historical exogenous time-series data. Based on these historical time-series and exogenous characteristics, a second forecast load time-series data for the forecast time range is obtained. Further, based on the first and second forecast load time-series data, the target forecast load time-series data for the forecast time range is obtained. In this disclosure, by acquiring the second forecast load time-series data for the forecast time range based on historical time-series and exogenous characteristics, and by obtaining the final target forecast load time-series data based on the first and second forecast load time-series data, load forecasting achieves a combination of features in the time-series dimension and features in the exogenous data dimension, improving the accuracy of load forecasting, making the load forecasting method more applicable and practical, and optimizing the load forecasting method.

[0103] To achieve the above objectives, this disclosure also proposes a training method for a load forecasting model, which can be combined with... Figure 3 To understand further, Figure 3This is a flowchart illustrating a training method for a load prediction model according to another embodiment of the present disclosure, as shown below. Figure 3 As shown, the method includes:

[0104] S301, acquire sample load time series data and sample exogenous time series data within the sample time range.

[0105] In this embodiment of the disclosure, load forecasting within a forecast time range can be performed based on a trained load forecasting model. In order to improve the model performance of the trained load forecasting model, a load forecasting model to be trained can be obtained, and the model can be trained based on relevant sample data.

[0106] In practice, the factors influencing load changes can include factors influencing time series and factors influencing external data. Therefore, the load prediction model to be trained can be based on load time series data and external time series data within a historical time range.

[0107] Optionally, based on a set sample time range, the corresponding load time series data within the sample time range can be obtained as sample load time series data for model training.

[0108] Correspondingly, exogenous time series data corresponding to the sample time range can also be obtained as exogenous time series data for model training.

[0109] Furthermore, based on the sample load time series data and the sample exogenous time series data, the sample data required for training the load prediction model to be trained is obtained.

[0110] S302, input the sample load time series data and the sample exogenous time series data into the load prediction model to be trained. The load prediction model obtains the first sample load change information of the sample load time series data under the influence of the sample exogenous time series data, and extracts the first sample feature of the first sample load change information.

[0111] In this embodiment of the disclosure, feature extraction can be performed on sample load time series data and sample exogenous time series data through the load prediction model to be trained.

[0112] Specifically, sample load time series data and sample exogenous time series data can be input into the load prediction model to be trained. Based on the load prediction model to be trained, the load change information generated by the sample load data in the sample load time series data under the influence of sample exogenous time series data is obtained, and it is determined as the first sample load change information.

[0113] Furthermore, based on the feature extraction layer of the load prediction model to be trained, the first sample load change information is extracted, and the first sample features corresponding to the first sample load change information are obtained from the extraction results of the feature extraction layer.

[0114] S303, based on the characteristics of the first sample, obtain the first load time series data of the sample time range under the influence of exogenous time series data of the sample.

[0115] In this embodiment of the disclosure, the load prediction model to be trained can perform load prediction based on the first sample features, wherein the first sample features can be understood as the change characteristics of the sample load data in the sample load time series data under the influence of the sample exogenous time series data.

[0116] Furthermore, load prediction can be performed based on the features of the first sample, thereby obtaining the load data that the load prediction model may reach within the sample time range under the influence of exogenous time series data obtained during the training process.

[0117] Specifically, the load data obtained by the load forecasting model based on the characteristics of the first sample can be determined as the first load time series data within the sample time range.

[0118] S304. Based on the first load time series data and the sample load time series data, adjust the load prediction model and return to continue training the adjusted load prediction model to obtain the trained target load prediction model.

[0119] In this embodiment of the disclosure, the first load time series data is obtained from the load prediction model to be trained, and is the load time series data that may be achieved based on the influence of the sample exogenous data within the sample time range.

[0120] Optionally, the load forecasting model can obtain the loss of load forecasting based on the exogenous data dimension based on the first load time series data and the sample load time series data, and the load forecasting model can be adjusted according to the loss.

[0121] Furthermore, the model can be trained again using the next sample of exogenous time-series data and the next sample of load time-series data until the training termination condition is met. Then, the model training can be terminated and the load prediction model obtained after training can be determined as the trained target load prediction model.

[0122] In some implementations, a corresponding training termination condition can be set based on the loss of the load forecasting model in the dimension of exogenous data. If the loss between the first load time series data obtained in a certain round of model training and the sample load time series data input to the model in that round meets the training termination condition, the model training of the load forecasting model can be terminated, and the load forecasting model after the end of that round of training can be used as the trained target load forecasting model.

[0123] In other implementations, a corresponding training termination condition can be set based on the number of training rounds of the load prediction model. The training rounds of the load prediction model can be monitored and recorded. If the total number of training rounds recorded after a certain round of model training meets the training termination condition, the training of the load prediction model can be terminated, and the load prediction model after the last round of training can be used as the trained target load prediction model.

[0124] The load forecasting model training method proposed in this disclosure acquires sample load time-series data and sample exogenous time-series data within a sample time range. Based on the load forecasting model to be trained, it acquires first sample load change information under the influence of sample exogenous data and extracts first sample features from the first sample load change information. Further, based on the first sample features, it acquires first load time-series data within the sample time range under the influence of sample exogenous time-series data. Based on the first load time-series data and the sample load time-series data input to the model, the load forecasting model is adjusted and optimized, and the adjusted load forecasting model is returned for further model training until the training is completed and a trained target load forecasting model is obtained. In this disclosure, load forecasting within a sample time range is performed based on first sample features, realizing load forecasting based on exogenous time-series data, optimizing the load forecasting model training method, improving the accuracy of the trained target load forecasting model for load forecasting, optimizing the applicability and practicality of the trained target load forecasting model, and optimizing the load forecasting method.

[0125] To better understand the training method for the load forecasting model proposed in the above embodiments, it can be combined with... Figure 4 , Figure 4 This is a flowchart illustrating a training method for a load prediction model according to another embodiment of the present disclosure, as shown below. Figure 4 As shown, the method includes:

[0126] S401, Based on the load forecasting model, obtain the sample time series characteristics of the sample load time series data, wherein the sample time series characteristics include at least one of the sample time series trend characteristics and sample seasonal characteristics of the sample load time series data.

[0127] In practice, factors affecting changes in electricity load data can include external factors, such as weather, as well as time-series factors, such as time or season.

[0128] Furthermore, the load prediction model to be trained can be trained by combining the sample time series features in the sample load time series data.

[0129] like Figure 5 As shown, sample load time series data within the sample time range, as well as sample exogenous time series data including sample observed variable time series data, sample static variable time series data, and sample covariate time series data within the sample time range, can be input into the load prediction model to be trained.

[0130] The time-series data for sample observation variables may include voltage time-series data, reactive power time-series data, and measured weather time-series data within the sample time range. The time-series data for sample static variables may include predicted weather time-series data and sample time-series data within the sample time range. The time-series data for sample variable variables may include time-series data corresponding to the installed capacity of motors and time-series data corresponding to bus IDs within the sample time range.

[0131] like Figure 5 As shown, the load forecasting model performs load forecasting within the sample time range based on the sample time series data and exogenous data within the input sample time range.

[0132] Optionally, such as Figure 5 The load values ​​shown are the output of the load forecasting model. The load values ​​can be understood as the load data at the corresponding sample timestamp within the sample time range obtained by the load forecasting model when performing load forecasting.

[0133] Alternatively, it can also include, for example Figure 5 The load confidence interval shown is the output of the load forecasting model. The load confidence interval is the load value interval and the corresponding confidence probability of the load value interval on the corresponding sample timestamp within the sample time range obtained by the load forecasting model when performing load forecasting.

[0134] For example, the load forecasting model performs load forecasting within the sample time range based on the sample load time series data input to the model, as well as the sample exogenous time series data including the sample observed variable time series data, the sample static variable time series data, and the sample covariate time series data within the sample time range. The forecasting result output by the load forecasting model can be a load value range of [0.2, 0.5] and the corresponding 80% confidence probability of the load value range.

[0135] Optionally, based on the load forecasting model, features can be extracted from the time-series data of the sample load, and the extracted features can be determined as the sample time-series features of the sample load time-series data.

[0136] Among them, the time-series trend features in the sample time-series data can be extracted based on the load forecasting model as the sample time-series trend features, and the seasonal features in the sample time-series data can also be extracted based on the load forecasting model as the sample seasonal features.

[0137] Furthermore, based on the sample time-series trend characteristics and sample seasonal characteristics, the sample time-series data are obtained as sample time-series characteristics.

[0138] It should be noted that the sample time series features can be sample time series trend features, sample seasonal features, or both sample time series trend features and sample seasonal features. No specific limitation is made here.

[0139] S402, based on the sample time series characteristics, obtain the second load time series data within the sample time range obtained from the intermediate layer of the load forecasting model.

[0140] In some implementations, load forecasting can be performed based on sample time-series characteristics through the intermediate layer of the load forecasting model, and the load data obtained from the intermediate layer of the load forecasting model based on sample time-series characteristics is determined as the second load time-series data.

[0141] In other implementations, load forecasting can be performed within the sample time range by using the intermediate layer of the load forecasting model, based on the sample time series characteristics and the characteristics corresponding to the sample exogenous time series data. The obtained forecasting results are then used as the second load time series data obtained by the intermediate layer of the load forecasting model.

[0142] Among them, the exogenous features of the sample corresponding to the exogenous time series data can be obtained.

[0143] In this embodiment of the disclosure, features can be extracted from the exogenous time series data of the sample based on the feature extraction layer of the load forecasting model, and the exogenous time series data of the sample can be processed by the corresponding feature extraction algorithm in the feature extraction layer.

[0144] Among them, the features corresponding to the exogenous time series data of the sample can be obtained from the results of the algorithm processing, and these features can be identified as exogenous features of the sample.

[0145] Furthermore, the second sample load change information of the sample load time series data is obtained under the combined influence of sample time series features and sample exogenous features, and the second sample features corresponding to the second sample load change information are extracted.

[0146] In this embodiment of the disclosure, the intermediate layer of the load forecasting model can perform load forecasting within a sample time range based on the acquired sample time series characteristics and sample exogenous characteristics, obtain the load time series data that may be reached within the sample time range under the combined influence of the sample time series characteristics and sample exogenous characteristics, and determine it as the second load time series data.

[0147] Among them, the intermediate layer of the load forecasting model can be used to obtain the load change information of the sample load data in the sample load time series data under the combined influence of the sample time series characteristics and the sample exogenous characteristics, and determine it as the second sample load change information.

[0148] Optionally, based on the feature extraction algorithm in the feature extraction layer of the load forecasting model, the load change information of the second sample is processed by the algorithm, and the second sample features corresponding to the load change information of the second sample obtained by the feature extraction layer of the load forecasting model are obtained according to the result of the algorithm processing.

[0149] It should be noted that the second sample feature can be a time-series trend feature, a seasonal feature, or both; no specific limitation is made here.

[0150] Furthermore, based on the characteristics of the second sample, the second load time series data within the sample time range are obtained.

[0151] In this embodiment of the disclosure, load forecasting can be performed within a sample time range based on the second sample features through the intermediate layer of the load forecasting model.

[0152] Optionally, the second sample features can be analyzed and predicted based on the intermediate layer of the relevant prediction in the load forecasting model, and the load data within the sample time range predicted by the load forecasting model based on the second sample features can be determined according to the output results of the intermediate layer of the relevant prediction.

[0153] Specifically, the load data within the sample time range predicted by the load forecasting model based on the characteristics of the second sample can be determined as the second load time series data within the sample time range.

[0154] It should be noted that the load forecasting model can predict the second load time series data within the sample time range based on the second sample features, which include the sample time series trend characteristics and the sample exogenous characteristics. Alternatively, it can predict the second load time series data within the sample time range based on the second sample features, which include the sample seasonal characteristics and the sample exogenous characteristics. No specific limitation is made here.

[0155] S403: Based on the first load time series data and the second load time series data, obtain the target load time series data output by the load forecasting model.

[0156] In this embodiment of the disclosure, the load prediction model to be trained can analyze the sample time range to obtain the sample time series information carried in the sample time range.

[0157] Furthermore, based on the sample time series information, the first load time series data and the second load time series data are integrated in the time series dimension, and the integrated time series data is determined as the target load time series data predicted by the load forecasting model.

[0158] S404: Based on the target load time series data and the sample load time series data, adjust the load prediction model and return to continue training the adjusted load prediction model to obtain a trained target load prediction model.

[0159] In this embodiment of the disclosure, the training loss of the load prediction model can be obtained based on the target load time series data and the sample load time series data.

[0160] Among these methods, the loss value can be calculated for the target load time series data and the sample load time series data based on the loss function in the relevant technology, and the training loss of the load prediction model can be determined based on the calculated loss between the target load time series data and the sample load time series data.

[0161] Furthermore, the model parameters of the load prediction model are adjusted based on the training loss, and the adjusted load prediction model is trained again using the next sample load time series data and the next sample exogenous time series data until the training is completed, and the target load prediction model is obtained.

[0162] In this embodiment of the disclosure, the training samples for training the load prediction model to be trained may include multiple sample load time series data and sample exogenous time series data. The training samples can be divided based on the time series information to obtain multiple batches of training samples. Each batch of training samples includes sample load time series data and sample exogenous time series data under the same time series information.

[0163] Furthermore, after adjusting and optimizing the load prediction model to be trained based on the training loss, the model can be trained again by obtaining the next sample load time series data and the next sample exogenous time series data from the next batch of samples from the training samples, until the training ends.

[0164] In some implementations, a corresponding training termination condition can be set based on the training loss of the load prediction model. If the training loss obtained in a certain round of model training meets the training termination condition, the training of the load prediction model can be terminated, and the load prediction model after the end of that round of training can be used as the trained target load prediction model.

[0165] In other implementations, a corresponding training termination condition can be set based on the number of training rounds of the load prediction model. The training rounds of the load prediction model can be monitored and recorded. If the total number of training rounds recorded after a certain round of model training meets the training termination condition, the training of the load prediction model can be terminated, and the load prediction model after the last round of training can be used as the trained target load prediction model.

[0166] like Figure 6 As shown, historical load time series data, historical observed variable time series data, historical predicted covariate time series data, and historical static variable time series data can be obtained within the sample time range. Furthermore, the obtained historical observed variable time series data, historical predicted covariate time series data, and historical static variable time series data within the sample time range are integrated to obtain the corresponding historical exogenous time series data within the sample time range.

[0167] Specifically, the historical load time series data and historical exogenous time series data within the obtained sample time range can be divided into datasets to obtain sample load time series data and sample exogenous time series data used for training the load prediction model to be trained, and test load time series data and test exogenous time series data used for testing the model performance of the trained target load prediction model within the sample time range.

[0168] Furthermore, the sample load time series data and the sample exogenous time series data are input into the load prediction model to be trained for model training, thereby obtaining the output results of the load prediction model during the model training process.

[0169] Accordingly, after the load forecasting model has been trained, the training effect of the trained load forecasting model can be tested by testing load time series data and testing exogenous time series data, so as to obtain the trained target load forecasting model.

[0170] The training method for the load forecasting model proposed in this disclosure involves acquiring the sample time-series features of the sample load time-series data and the sample exogenous time-series features. Based on these features, load forecasting is performed within the sample time range to obtain second load time-series data for that range. Then, based on the first and second load time-series data, the target load time-series data obtained from the load forecasting model within the sample time range is derived. Further, based on the target load time-series data within the sample time range and the sample load time-series data input to the load forecasting model to be trained, the training loss of the load forecasting model is obtained. The model parameters are adjusted based on the obtained training loss. After the parameter adjustment is complete, the model is trained again using the next sample load time-series data and the next sample exogenous time-series data until the training is complete and the trained target load forecasting model is obtained. In this disclosure, a load forecasting model is trained based on sample load time-series data and sample exogenous time-series data within a sample time range. This enables the load forecasting model to support the joint modeling of multiple variables and learn the correlation features between changes in exogenous data and load data. Consequently, the load forecasting model can obtain second load time-series data within the sample time range based on sample time-series features and sample exogenous features. Based on the first load time-series data and the second load time-series data, the final load forecasting model obtains the target load time-series data for load forecasting within the sample time range. This achieves load forecasting that combines features in the time-series dimension with features in the exogenous data dimension, optimizes the training method of the load forecasting model, improves the accuracy of the trained target load forecasting model for load forecasting, optimizes the applicability and practicality of the trained target load forecasting model, and optimizes the load forecasting method.

[0171] In the above embodiments, the acquisition of target sample load time series data can also be combined with Figure 7 To understand further, Figure 7 This is a flowchart illustrating a training method for a load prediction model according to another embodiment of the present disclosure, as shown below. Figure 7 As shown, the method includes:

[0172] S701, acquire first load time series data from at least one stack output of a plurality of cascaded stacks of the load forecasting model, and second load time series data from the remaining stack outputs other than at least one stack.

[0173] In this embodiment of the disclosure, a load prediction model to be trained can be obtained based on a set structure. For example, a load prediction model to be trained can be constructed based on the basic model structure of a deep double residual network. On the basis of the deep double residual network model structure, a stack for feature extraction based on exogenous time series data and load prediction is added, and the input data of exogenous data is added to the input data of each stack, so that the load prediction model can learn the correlation features between exogenous time series data and load time series data.

[0174] Optionally, the load forecasting model may include multiple cascaded stacks. In this scenario, the order of the stacks in the load forecasting model may affect the stack input data of the stacks.

[0175] like Figure 8 As shown, time series data X and time series data Y are sample time series data used for training the load prediction model. Time series data Y can be exogenous time series data within the sample time range, while time series data X includes sample load time series data within the sample time range, such as... Figure 8 As shown, sample load time series data and sample exogenous time series data can be used as input data for the load forecasting model.

[0176] In some implementations, if the stack is the first stack of the load forecasting model, the sample load time series data and the sample exogenous time series data are used as the stack input data for the first stack.

[0177] like Figure 8 As shown, stack 1 is the first stack in the load forecasting model. In this scenario, the input data of the load forecasting model can be used as the input data of stack 1.

[0178] This can be understood as using the sample load time series data and the sample exogenous time series data of the input model as the stack input data of stack 1.

[0179] In some other implementations, if the stack is the (i+1)th stack of the load forecasting model, the stack residual data of the i-th stack is obtained, and the stack residual data of the i-th stack and the sample exogenous data are used as the stack input data of the (i+1)-th stack, where i is a positive integer greater than or equal to 1.

[0180] In this context, non-first stacks in the load forecasting model can be identified as the (i+1)th stack, where i is an integer greater than or equal to 1.

[0181] like Figure 8 As shown, stack 2 and stack 3 are the (i+1)th stacks in the load forecasting model, where the value of i for stack 2 is 1 and the value of i for stack 3 is 2.

[0182] In this embodiment of the disclosure, different stacks can achieve feature extraction of different dimensions, thereby enabling load prediction based on different feature dimensions.

[0183] Therefore, after the stack performs feature extraction and load prediction on the input data, there may be some data that is not extracted. This part of the data can be identified as stack residual data.

[0184] For the (i+1)th stack, its stack input data may include the stack residual data output by the i-th stack of the previous stack in its cascading relationship. For example... Figure 8 As shown, the stack input data of stack 2 may include the stack residual data output by stack 1, and the stack input data of stack 3 may include the stack residual data output by stack 2.

[0185] In this embodiment of the disclosure, in order to enable the load forecasting model to learn the correlation characteristics between exogenous time series data and load time series data, the input of sample exogenous time series data can be added to the stack input data of the stack in the load forecasting model.

[0186] In this scenario, the stack input data of the (i+1)th stack can include the stack residual data output by the ith stack and sampled exogenous time-series data. For example... Figure 8 As shown, the stack input data of stack 2 may include the stack residual data output by stack 1 and the sample exogenous time series data of the input load prediction model, and the stack input data of stack 3 may include the stack residual data output by stack 2 and the sample exogenous time series data of the input load prediction model.

[0187] In this embodiment of the disclosure, different stacks can perform load prediction based on features extracted in different dimensions, thereby obtaining predicted load time series data under different feature dimensions, and then obtaining the target load time series data output by the load prediction model.

[0188] In this scenario, the component load time series data output by each block in each of the multiple cascaded blocks in the load forecasting model can be obtained. Each block can include at least two output branches, and the component load time series data output by the corresponding block can be obtained from the output results of the output branches that output the component load time series data.

[0189] Furthermore, based on the component load timing data output by each block, the load timing data output by the stack is obtained.

[0190] like Figure 8As shown, the residual data of block1 and the sample exogenous time series data input to block2 can be mixed in the time series data connection layer (cat layer) set in block2 to obtain mixed time series data of residual data of block1 and sample exogenous time series data. The obtained mixed time series data is then input into the fully connected layer in block2. After processing by the algorithm based on the fully connected layer, the processing output is divided into two paths and input into the linear layers on the two branches respectively.

[0191] Furthermore, based on the calculation and processing of linear functions in the linear layer, the corresponding output results are obtained, and the corresponding output results are input into g respectively. b (θ b ) layer and g f (θ f )layer.

[0192] Among them, g b (θ b The ) layer is used to extract sample temporal features and exogenous features from mixed time series data. f (θ f The layer is used for load prediction based on the sample time-series features and sample exogenous features in the extracted mixed time-series data.

[0193] Furthermore, from g b (θ b The temporal and exogenous features of the samples extracted by block2 can be obtained from the output of layer ) and g. f (θ f The output of the layer yields component load time series data obtained by block2 based on the extracted sample time series features and sample exogenous features for load prediction.

[0194] In this embodiment of the disclosure, there is an order among multiple cascaded blocks in the stack. In this scenario, the order of the blocks in the stack may affect the input data of the blocks.

[0195] In some implementations, if the block is the first block in the stack, the stack input data is input into the first block to obtain the first component load timing data output by the first block.

[0196] Among them, multiple cascaded blocks in the stack can be sorted from top to bottom. If a block is the first block in the sort, the stack input data of its own stack can be used as the input data of that block.

[0197] like Figure 8As shown, block1 is the first block in stack 2, so the stack input data of stack 2 can be used as the input data of block1. Figure 8 As can be seen, the stack input data of stack 2 includes the stack residual data output by stack 1 and the sample exogenous time series data. Therefore, the stack input data including the stack residual data output by stack 1 and the sample exogenous time series data can be used as the input data of block 1.

[0198] Among them, stack residual data can be understood as a mixture of sample load time series data and sample exogenous time series data remaining after feature extraction in a certain dimension.

[0199] In this embodiment of the disclosure, the block is provided with a corresponding layer for feature extraction and a corresponding layer for load prediction. Therefore, the block can extract features from the input data and perform load prediction based on the extracted features, thereby obtaining component load time series data obtained by load prediction based on the extracted feature dimensions.

[0200] Specifically, the first block of the sample load time series data and the first block of the sample exogenous time series data can be extracted from the first block.

[0201] Optionally, the time series data of the sample load included in the input data of the first block can be extracted by the corresponding layer of feature extraction set in the first block of the stack, so as to obtain the time series features of the sample load time series data extracted in the first block, and determine them as the first block time series features extracted in the first block.

[0202] Correspondingly, the feature extraction layer set in the first block can be used to extract features from the sample exogenous time series data included in the input data of the first block, thereby obtaining the exogenous data features of the sample exogenous time series data extracted in the first block, and determining them as the first block exogenous features extracted in the first block.

[0203] In this scenario, the first block can perform load prediction within the sample time range based on the extracted first block time series features and the first block exogenous features. Specifically, the first component load time series data within the sample time range obtained by the first block based on the first block time series features and the first block exogenous features can be obtained from the output branch of the first block.

[0204] like Figure 8As shown, block1 is the first block in stack 2. Block1 can extract temporal features and exogenous features from the input data, thereby obtaining the first block temporal features and the first block exogenous features extracted by block1.

[0205] Furthermore, block1 can perform load prediction within the sample time range based on the extracted first block time series features and the first block exogenous features, and obtain the first component load time series data predicted by block1 from the output of block1.

[0206] like Figure 8 As shown, block1 includes two output branches, namely output branch s1 and output branch s2. Output branch s2 is the output branch of the output component load time series data. Therefore, the first component load time series data obtained by block1 through feature extraction and load prediction of its input data can be obtained from the output of output branch s2.

[0207] In some other implementations, if the block is the (j+1)th block in the stack, the residual data of the jth block is obtained, and the residual data of the jth block and the sample exogenous time series data are input into the (j+1)th block to obtain the second component load time series data output by the (j+1)th block, where j is a positive integer greater than or equal to 1.

[0208] In this embodiment of the disclosure, after the block extracts temporal features from the input data, there may be unextracted temporal features in the remaining data. Therefore, for the (j+1)th block that is not the first block, the remaining residual data after the temporal features were extracted from the previous block can continue to be extracted for temporal features and exogenous features.

[0209] In this scenario, the input data for the (j+1)th block may include the residual data output after the temporal features are extracted from the j-th block, as well as the exogenous temporal data of the samples input to the (j+1)th block.

[0210] The residual data output after extracting time-series features from the j-th block can be understood as a mixture of the remaining sample load time-series data and the sample exogenous time-series data after extracting time-series features from the j-th block.

[0211] like Figure 8 As shown, the j-th block corresponding to block2 is... Figure 8As shown in block1, the input data of block2 can include the residual data output by block1. In this scenario, the mixed time series data of sample load time series data and sample exogenous time series data remaining after the extraction of time series features and exogenous features of block1 can be obtained from the output branch of block1.

[0212] Depend on Figure 8 It can be seen that the first block time series features and the first block exogenous features extracted by block1 can be obtained from the s1 output branch of the two output branches of block1. Furthermore, these can be processed with the mixed time series data of the sample load time series data and the sample exogenous time series data of input block1. The first block time series features and the first block exogenous features extracted by block1 are deleted from the mixed time series data of the sample load time series data and the sample exogenous time series data of input block1. The mixed time series data of the sample load time series data and the sample exogenous time series data obtained after deletion is the residual data remaining after the time series feature extraction of block1.

[0213] In this scenario, in order for block2 to learn the exogenous features in the sample exogenous time series data more completely, the sample exogenous time series data of the input model and the residual data of block1 can be used as the input data of block2.

[0214] Furthermore, the second block temporal features are extracted from the residual data of the j-th block through the (j+1)-th block, and the second block exogenous features are extracted from the sample exogenous temporal data.

[0215] Optionally, the time series features extracted from the mixed time series data of sample load time series data and sample exogenous time series data included in the residual data of the output of the jth block in the input data of the j+1th block can be extracted through the corresponding layer of feature extraction set in the j+1th block, thereby obtaining the time series features extracted from the j+1th block and determining them as the second block time series features extracted from the j+1th block.

[0216] Correspondingly, the feature extraction layer set in the (j+1)th block can be used to extract features from the sample exogenous time series data included in the input data of the (j+1)th block, thereby obtaining the exogenous data features of the sample exogenous time series data extracted in the (j+1)th block, and determining them as the second block exogenous features extracted in the (j+1)th block.

[0217] like Figure 8As shown, block2 is the (j+1)th block in stack 2, where j is 1. (From...) Figure 8 It can be seen that the residual data after time-series feature extraction in block1 and the sample exogenous time-series data can be used as the input data for block2.

[0218] Optionally, temporal and exogenous features can be extracted from the input data through block2. The temporal features extracted by block2 can be determined as the second block temporal features of block2, and the exogenous features extracted by block2 can be determined as the second block exogenous features of block2.

[0219] Furthermore, in the output branch of the (j+1)th block, the second component load time series data within the sample time range obtained by the (j+1)th block based on the time series characteristics of the second block and the exogenous characteristics of the second block are acquired.

[0220] In this scenario, the (j+1)th block can perform load prediction within the sample time range based on the extracted second block time series features and the second block exogenous features. Specifically, the second component load time series data within the sample time range obtained by the (j+1)th block based on the second block time series features and the second block exogenous features can be obtained from the output branch of the (j+1)th block.

[0221] like Figure 8 As shown, block2 includes two output branches, namely output branch s3 and output branch s4. Output branch s4 is the output branch for the component load time series data. Therefore, the second component load time series data obtained by feature extraction and load prediction of the input data of block2 can be obtained from the output results of output branch s4.

[0222] Furthermore, based on the component load timing data of each block output in each stack, first load timing data of at least one stack output and second load timing data of the remaining stack outputs other than at least one stack are obtained.

[0223] Optionally, the component load timing data output from each block in the stack can be integrated through a corresponding layer set in the stack to integrate the component load timing data, thereby obtaining the load timing data output by the stack.

[0224] Furthermore, the stack in the load forecasting model can extract features from the input time-series data in different dimensions, thereby obtaining load time-series data for prediction under different feature dimensions.

[0225] Among them, the load data obtained by the stack based on the load prediction within the sample time range based on the sample exogenous characteristics can be determined as the first load time series data obtained by the stack.

[0226] Optionally, the component load time series data obtained from load prediction within the sample time range based on sample exogenous features of each block output in the stack can be integrated to obtain the first load time series data of the stack output.

[0227] Accordingly, the load data obtained by the stack based on the combination of sample time-series features and sample exogenous features within the sample time range can be identified as the second load time-series data obtained by the stack.

[0228] Optionally, the component load time series data obtained from the load prediction within the sample time range based on the combination of sample time series features and sample exogenous features of each block output in the stack can be integrated to obtain the second load time series data output by the stack.

[0229] like Figure 8 As shown, if stack 1 is configured to perform load prediction within a sample time range based on the sample time series characteristics, including the sample time series trend characteristics and the sample exogenous characteristics, then stack 1 can be identified as a time series trend stack.

[0230] Stack 2 can predict the load within a time range of the sample based on the seasonal characteristics and exogenous characteristics of the sample time series features. Therefore, Stack 2 can be identified as a seasonal stack.

[0231] Stack 3 can predict the load over a time range of samples based on the exogenous characteristics of the samples, so stack 3 can be identified as an exogenous stack.

[0232] Depend on Figure 8 It can be seen that sample load time series data and sample exogenous time series data can be input into stack 1trend. The blocks in the trend stack extract sample time series trend features from sample load time series data and sample exogenous features from sample exogenous time series data. Based on the combination of sample time series trend features and sample exogenous features, load prediction within the sample time period is performed, resulting in component load time series data output by each block in the trend stack.

[0233] Optionally, the time-series trend features of the sample can be extracted from the mixed data of sample load time-series data and sample exogenous time-series data based on the formula shown below:

[0234]

[0235] In the above formula, θ is the coefficient, and i is the degree of t. Here, s represents the time-series trend features extracted from the trend stack, b represents the index of the block in the trend stack, l represents the layer number of the trend stack in the load forecasting model, b,i represents the i-th term in the formula, and t represents time. Tθ represents the summation from i = 0 to i = p, where p is the highest order. s,l for The matrix representation of .

[0236] Furthermore, based on the integration layer of component load time series data in the trend stack, the component load time series data output by each block are integrated to obtain the second load time series data output by the trend stack, which is based on the combination of sample time series trend features and sample exogenous features to predict the load within the sample time period.

[0237] Accordingly, such as Figure 8 As shown, the mixed time series data of sample load and sample exogenous time series data remaining after extracting the time series trend features from the stack residual data of the trend stack, as well as the sample exogenous time series data of the input model, can be used as the input data of the 2seasonality stack. The blocks in the seasonality stack extract the sample seasonal features from the stack residual data of the trend stack and the sample exogenous features from the sample exogenous time series data. Based on the combination of sample seasonal features and sample exogenous features, load prediction within the sample time period is performed to obtain the component load time series data output by each block in the seasonality stack.

[0238] Optionally, seasonal features of the sample can be extracted from the mixed data of sample load time series data and sample exogenous time series data based on the formula shown below:

[0239]

[0240] In the above formula, θ represents the coefficients of the trigonometric functions. The seasonality features extracted from the seasonality stack are: s is the exponent of the trend stack, b is the exponent of the block in the trend stack, l is the layer identifier of the trend stack in the load forecasting model, i represents the number of terms, b and i represent the i-th term of the formula, t is time, and H is the sequence length of the second load time series data predicted by the seasonality stack.

[0241] Furthermore, based on the integration layer of component load time series data in the seasonality stack, the component load time series data output by each block are integrated to obtain the second load time series data output by the seasonality stack, which is based on the combination of sample seasonality features and sample exogenous features to predict the load within the sample time period.

[0242] Furthermore, by Figure 8 It can be seen that the mixed time series data of sample loads and sample exogenous time series data remaining after seasonality features are extracted from the stack residual data of the seasonality stack, as well as the sample exogenous time series data of the input model, can be used as the input data of the stack 3exogenous stack.

[0243] In this scenario, the residual data of the seasonality stack may contain only a mixture of the input data of the trend stack and the seasonality stack, which are the remaining sample exogenous time series data after the exogenous features of the trend stack and the seasonality stack have been extracted. In order for the exogenous stack to learn the exogenous features in the sample exogenous data more completely, the sample exogenous time series data of the input model and the residual data of the seasonality stack can be used as the stack input data of the exogenous stack.

[0244] Optionally, the internal structure of an exogenous stack can be constructed based on a TCN network to enable the extraction of exogenous features from samples and the prediction of first load time series data based on these exogenous features.

[0245] Furthermore, the exogenous stack extracts exogenous features from the mixed time series data of the seasonality stack and the sample exogenous time series data, and performs load prediction within the sample time based on the exogenous features, thereby obtaining the first load time series data output by the exogenous stack based on the sample exogenous features for load prediction within the sample time.

[0246] In this embodiment of the disclosure, the acquisition of stack residual data in the load forecasting model can be understood with reference to the following example:

[0247] Furthermore, the stack residual data of the stack to which the block belongs can be obtained through the residual data output by the block. Specifically, the blocks can be sorted in the stack, the last block in the stack can be obtained, and the residual data output by the last block can be used as the stack residual data of its respective stack.

[0248] Optionally, the residual data of the (j+1)th block can be obtained based on the extracted temporal features of the second block and the exogenous features of the second block, the residual data of the jth block input to the (j+1)th block, and the mixed temporal data of the sample exogenous temporal data.

[0249] Specifically, the time-series features and exogenous features extracted from the output branch of the j-th block can be obtained and compared with the mixed time-series data of the sample load time-series data and sample exogenous time-series data input to the j-th block. The time-series features and exogenous features extracted from the j-th block are then deleted from the mixed time-series data of the sample load time-series data and sample exogenous time-series data input to the j-th block, thus obtaining the mixed time-series data of the sample load time-series data and sample exogenous time-series data remaining after deleting the time-series features and exogenous features extracted from the j-th block.

[0250] Furthermore, the mixed time series data of sample load time series data and sample exogenous time series data obtained after the deletion process is the residual data remaining after the time series features and exogenous features are extracted from the j-th block.

[0251] Specifically, the residual connection between the input and output of the j-th block can be used to obtain the mixed time series data of the sample load time series data and the sample exogenous time series data of the j-th block at the output of the j-th block. Then, it is processed with the time series features and exogenous features output by the j-th block to obtain the residual data corresponding to the j-th block.

[0252] like Figure 8 As shown, the second block temporal features and the second block exogenous features extracted by block2 can be obtained from the output branch s3 of block2. Through the residual connection from the input end to the output end of block2, the mixed temporal data of the residual data of block1 input to block2 and the sample exogenous temporal data can be obtained, and then the residual data after the temporal features and exogenous features of block2 are obtained.

[0253] In some implementations, if the (j+1)th block is not the last block in the stack sorted from top to bottom, the residual data of the (j+1)th block is determined as the input data of the (j+2)th block.

[0254] In the scenario where the (j+1)th block is not the last block in the stack after being sorted from top to bottom, there is a (j+2)th block cascaded after the (j+1)th block. In this scenario, the residual data corresponding to the (j+1)th block can be used as the input data for the (j+2)th block.

[0255] like Figure 8 As shown, block2 is the (j+1)th block in stack 2 that is not the last one, and is composed of... Figure 8 As can be seen, there is a block that is cascaded after block2. In this scenario, the residual data corresponding to block2 can be used as the input data for the next cascaded block3 (not shown in the figure).

[0256] In some other implementations, if the (j+1)th block is the last block of the stack sorted from top to bottom, the residual data of the (j+1)th block is determined as the stack residual data of the stack.

[0257] In the scenario where the (j+1)th block is the last block in its stack after being sorted from top to bottom, there is no next-level block after the (j+1)th block. In this scenario, the residual data corresponding to the (j+1)th block can be determined as the stack residual data of the stack to which the (j+1)th block belongs.

[0258] like Figure 8 As shown, blockN is the last block in stack 2 ordered from top to bottom. In this scenario, there is no next-level block after blockN, so the residual data of blockN can be used as the stack residual data of stack 2.

[0259] Specifically, by using the residual connection between the input and output of blockN, a mixed time series data of sample load time series data and sample exogenous time series data of input blockN can be obtained. This mixed time series data is then compared with the time series features and exogenous features extracted by blockN. The time series features and exogenous features extracted by blockN are then deleted from the mixed time series data of sample load time series data and sample exogenous time series data of input blockN, thereby obtaining the residual data of blockN after deletion processing, which is then identified as the stack residual data of stack 2.

[0260] S702, integrate the first load time series data and the second load time series data in terms of time series dimension to obtain the integrated target load time series data.

[0261] In this embodiment of the present disclosure, sample time series information within a sample time range can be obtained based on a load forecasting model, and the first load time series data and the second load time series data can be integrated in the time series dimension based on the sample time series information to obtain the integrated target load time series data.

[0262] Optionally, time-series information within the sample time range can be acquired, and the corresponding first load data and second load data on the same time-series information can be integrated to obtain the target load data on the same time-series information.

[0263] Furthermore, the target load data corresponding to all time series information within the sample time range are integrated based on the time series within the sample time range to obtain the final load data obtained by the load forecasting model based on the sample exogenous time series data and the sample load time series data within the sample time range. This final load data is then determined as the target load time series data obtained by the load forecasting model for load forecasting within the sample time range.

[0264] The training method for the load forecasting model proposed in this disclosure uses first or second load time-series data output from multiple cascaded stacks in the load forecasting model. Each stack's first or second load time-series data can be obtained by integrating component load time-series data output from multiple cascaded blocks within the stack. Further, based on the first and second load time-series data, the target load forecasting data within the final sample time range obtained by the load forecasting model is acquired. This disclosure, by adding stacks corresponding to exogenous time-series data and adding inputs of sample exogenous time-series data to the input data of each stack, enables the load forecasting model to more completely learn the correlation features between sample exogenous time-series data and sample load time-series data. This improves the learning rate and interpretability of the load forecasting model, optimizes the training method and training effect, and allows the trained target load forecasting model to combine time-series features and exogenous features for load forecasting, improving the prediction accuracy of the target load forecasting model, optimizing the applicability and practicality of the trained target load forecasting model, and ultimately optimizing the load forecasting method.

[0265] To better understand the above embodiments, this disclosure also proposes a load timing prediction method, which can be combined with... Figure 9 To understand further, Figure 9 This is a flowchart illustrating a load timing prediction method according to an embodiment of the present disclosure, as shown below. Figure 9As shown, the method includes:

[0266] S901, acquire the historical load time series data to be predicted, as well as the predictable exogenous time series data within the prediction time range.

[0267] In this embodiment of the disclosure, load forecasting within a set forecast time range can be performed using a trained target load forecasting model, wherein the target load forecasting model is obtained based on the training method of the load forecasting model proposed in the above embodiment.

[0268] Among these, historical load time-series data to be predicted, as well as predictable exogenous time-series data within the prediction time range, can be obtained from storage locations in relevant systems.

[0269] Optionally, the predictable exogenous time series data may include predictable observed variable time series data, predictable predictable covariate time series data, and predictable static variable time series data within the prediction time range.

[0270] Furthermore, the time-series data of predictable observed variables, predictable covariates, and predictable static variables within the forecast time range can be integrated along the time-series dimension to obtain the predictable exogenous time-series data required for load forecasting within the forecast time range.

[0271] S902: Input historical load time series data and predictable exogenous time series data into the trained target load prediction model, and obtain the target predicted load time series data within the prediction time range from the output of the target load prediction model, based on historical load time series data and predictable exogenous time series data.

[0272] In this embodiment of the disclosure, the acquired historical load time series data and predictable exogenous time series data can be input into the trained target load prediction model. The target load prediction model extracts features of a set dimension from the input historical load time series data and predictable exogenous time series data, and performs load prediction within the prediction time range based on the extracted features.

[0273] Furthermore, from the output of the target load forecasting model, load time series data within the forecast time range obtained by the target load forecasting model based on historical load time series data and predictable exogenous time series data are obtained, and these are determined as the target forecast load time series data output by the target load forecasting model.

[0274] Optionally, the effectiveness of load forecasting based on the trained target load forecasting model can be combined with... Figure 10 Understanding, such as Figure 10 As shown:

[0275] in, Figure 10 Image number 1 is a comparison chart between the target predicted load time series data output by the target load prediction model based on the input historical load time series data and predictable exogenous time series data, and the actual load time series data reached within the prediction time series.

[0276] Image number 2 is a comparison chart between the time-series trend predicted load time-series data obtained by predicting the load within the prediction time range based on the time-series trend characteristics of the input historical load time-series data in the target load prediction model, and the time-series trend load time-series data actually reached within the prediction time range only under the influence of the time-series trend characteristics.

[0277] Image number 3 is a comparison chart between seasonal load time series data obtained by predicting the load within the prediction time range based on the seasonal characteristics of the input historical load time series data in the target load prediction model, and seasonal load time series data actually reached within the prediction time range only under the influence of seasonal characteristics.

[0278] Image number 4 is a comparison chart of the predicted load time series data output by the target load prediction model based on historical load time series data and historical exogenous time series data within the input historical time range, and the historical load time series data input to the model.

[0279] Image number 5 is a comparison chart of the historical load time series data obtained by the target load prediction model based on the time series trend characteristics in the input historical load time series data, and the historical load time series data input to the model.

[0280] Image number 6 is a comparison chart of historical seasonal load forecast time series data obtained by the target load forecasting model based on the seasonal characteristics in the input historical load time series data, and the historical load time series data input to the model.

[0281] Depend on Figure 10 It can be seen that by learning the correlation characteristics between exogenous time series data and load time series data within the same time range, the prediction accuracy of the target load forecasting model for load forecasting is effectively improved.

[0282] The load time-series forecasting method proposed in this disclosure acquires historical load time-series data to be predicted, as well as predictable exogenous time-series data within the prediction time range, and inputs them into a trained target load forecasting model. The target predicted load time-series data within the prediction time range is obtained through the output of the target load forecasting model. In this disclosure, load forecasting is performed using a trained target load forecasting model, achieving load forecasting by combining exogenous time-series data and historical load time-series data, effectively improving the accuracy of load forecasting and optimizing the load forecasting method.

[0283] Corresponding to the load forecasting methods proposed in the above embodiments, an embodiment of this disclosure also proposes a load forecasting device. Since the load forecasting device proposed in this disclosure corresponds to the load forecasting methods proposed in the above embodiments, the implementation methods of the above load forecasting methods are also applicable to the load forecasting device proposed in this disclosure, and will not be described in detail in the following embodiments.

[0284] Figure 11 This is a schematic diagram of the structure of a load forecasting device according to an embodiment of the present disclosure, as shown below. Figure 11 As shown, the load forecasting device 1100 includes a first acquisition module 111, a first extraction module 112, a first forecasting module 113, and a second forecasting module 114, wherein:

[0285] The first acquisition module 111 is used to acquire historical load time series data, historical exogenous time series data, and predictable exogenous time series data within the prediction time range;

[0286] The first extraction module 112 is used to obtain the first load change information of historical load time series data under the influence of historical exogenous time series data, and extract the first feature corresponding to the first load change information;

[0287] The first prediction module 113 is used to obtain, based on the first feature, the first predicted load time series data within the prediction time range under the influence of predictable exogenous time series data;

[0288] The second prediction module 114 is used to obtain the target predicted load time series data within the prediction time range based on the first predicted load time series data.

[0289] In this embodiment of the disclosure, the second prediction module 114 is further configured to: obtain historical time-series features corresponding to historical load time-series data, wherein the historical time-series features include at least one of historical time-series trend features and historical seasonal features; obtain second predicted load time-series data within the prediction time range based on the historical time-series features; and obtain target predicted load time-series data based on the first predicted load time-series data and the second predicted load time-series data.

[0290] In this embodiment of the disclosure, the second prediction module 114 is further configured to: acquire historical exogenous features corresponding to historical exogenous time series data; acquire second load change information of historical load time series data under the combined influence of historical time series features and historical exogenous features, and extract second features corresponding to the second load change information; and acquire second predicted load time series data within the prediction time range based on the second features.

[0291] In this embodiment of the disclosure, the second prediction module 114 is further configured to: obtain prediction time series features corresponding to the prediction time range, wherein the prediction time series features include at least one of prediction time series trend features and prediction seasonality features corresponding to the prediction time range; obtain predictable exogenous features within the prediction time range from predictable exogenous time series data; and obtain second prediction load time series data within the prediction time range under the combined influence of the prediction time series features and predictable exogenous features, based on the second feature.

[0292] In this embodiment of the present disclosure, the first acquisition module 111 is further configured to: acquire time series data of observed variables, time series data of predicted covariates, and time series data of static variables within the prediction time range; and integrate the time series data of observed variables, time series data of predicted exogenous variables, and time series data of static variables in the time series dimension based on the time series information within the prediction time range to obtain predictable exogenous time series data within the prediction time range.

[0293] In this embodiment of the present disclosure, the second prediction module 114 is further configured to: integrate the first predicted load time series data and the second predicted load time series data according to the time series to obtain integrated load time series data, which is used as the target predicted load time series data.

[0294] The load forecasting device proposed in this disclosure acquires historical time-series characteristics of historical load time-series data and historical exogenous characteristics of historical exogenous time-series data. Based on these historical time-series and exogenous characteristics, it acquires second predicted load time-series data for a prediction time range. Further, based on the first and second predicted load time-series data, it obtains target predicted load time-series data for the prediction time range. In this disclosure, by acquiring second predicted load time-series data for the prediction time range based on historical time-series and exogenous characteristics, and by obtaining the final target predicted load time-series data based on the first and second predicted load time-series data, it achieves load forecasting that combines features in the time-series dimension with features in the exogenous data dimension. This improves the accuracy of load forecasting, makes the load forecasting method more applicable and practical, and optimizes the load forecasting method.

[0295] Corresponding to the training methods of the load forecasting models proposed in the above embodiments, an embodiment of this disclosure also proposes a training device for the load forecasting model. Since the training device for the load forecasting model proposed in this disclosure corresponds to the training methods of the load forecasting models proposed in the above embodiments, the implementation methods of the above-mentioned load forecasting models are also applicable to the training device for the load forecasting model proposed in this disclosure, and will not be described in detail in the following embodiments.

[0296] Figure 12 This is a schematic diagram of the structure of a training device for a load prediction model according to an embodiment of the present disclosure, as shown below. Figure 12 As shown, the training device 1200 for the load prediction model includes a second acquisition module 121, a second extraction module 122, a third prediction module 123, a training module 124, and a stack extraction prediction module 125, wherein:

[0297] The second acquisition module 121 is used to acquire sample load time series data and sample exogenous time series data within the sample time range;

[0298] The second extraction module 122 is used to input the sample load time series data and the sample exogenous time series data into the load prediction model to be trained. The load prediction model obtains the first sample load change information of the sample load time series data under the influence of the sample exogenous time series data, and extracts the first sample feature of the first sample load change information.

[0299] The third prediction module 123 is used to obtain the first load time series data of the sample time range under the influence of the sample exogenous time series data based on the characteristics of the first sample;

[0300] The training module 124 is used to adjust the load prediction model based on the first load time series data and the sample load time series data, and return to continue training the adjusted load prediction model to obtain the trained target load prediction model.

[0301] In this embodiment of the present disclosure, the training module 124 is further configured to: obtain sample time-series features of sample load time-series data based on the load forecasting model, wherein the sample time-series features include at least one of sample time-series trend features and sample seasonality features of the sample load time-series data; obtain second load time-series data within the sample time range obtained by the intermediate layer of the load forecasting model based on the sample time-series features; obtain target load time-series data output by the load forecasting model according to the first load time-series data and the second load time-series data; adjust the load forecasting model according to the target load time-series data and the sample load time-series data, and return to continue training the adjusted load forecasting model to obtain a trained target load forecasting model.

[0302] In this embodiment of the disclosure, the training module 124 is further configured to: acquire sample exogenous features corresponding to sample exogenous time series data; acquire second sample load change information of sample load time series data under the combined influence of sample time series features and sample exogenous features, and extract second sample features corresponding to the second sample load change information; and acquire second load time series data within the sample time range based on the second sample features.

[0303] In this embodiment of the present disclosure, the training module 124 is further configured to: acquire first load time-series data output from at least one stack in a plurality of cascaded stacks of the load prediction model, and second load time-series data output from the remaining stacks other than at least one stack; and integrate the first load time-series data and the second load time-series data in a time-series dimension to obtain integrated target load time-series data.

[0304] In this embodiment of the present disclosure, the apparatus further includes a stack extraction and prediction module 125, configured to: if the stack is the first stack of the load prediction model, use the sample load time series data and the sample exogenous time series data as the stack input data of the first stack; if the stack is the (i+1)th stack of the load prediction model, obtain the stack residual data of the i-th stack, and use the stack residual data of the i-th stack and the sample exogenous data as the stack input data of the (i+1)th stack, where i is a positive integer greater than or equal to 1.

[0305] In this embodiment of the present disclosure, the stack extraction prediction module 125 is further configured to: acquire component load time series data output by each block in the multiple cascaded blocks of each stack in the load prediction model; and acquire first load time series data output by at least one stack and second load time series data output by the remaining stacks other than at least one stack based on the component load time series data output by each block in each stack.

[0306] In this embodiment of the disclosure, the stack extraction prediction module 125 is further configured to: if the block is the first block in the stack, input the stack input data into the first block to obtain the first component load time series data output by the first block; if the block is the (j+1)th block in the stack, obtain the residual data of the jth block, and input the residual data of the jth block and the sample exogenous time series data into the (j+1)th block to obtain the second component load time series data output by the (j+1)th block, where j is a positive integer greater than or equal to 1.

[0307] In this embodiment of the disclosure, the stack extraction prediction module 125 is further configured to: extract the first block time series features in the sample load time series data and the first block exogenous features in the sample exogenous time series data through the first block; and obtain the first component load time series data within the sample time range obtained by the first block based on the first block time series features and the first block exogenous features from the output branch of the first block.

[0308] In this embodiment of the disclosure, the stack extraction prediction module 125 is further configured to: extract the second block time series features from the residual data of the j-th block through the (j+1)-th block, and extract the second block exogenous features from the sample exogenous time series data; and obtain the second component load time series data within the sample time range obtained by the (j+1)-th block based on the second block time series features and the second block exogenous features from the output branch of the (j+1)-th block.

[0309] In this embodiment of the present disclosure, the stack extraction prediction module 125 is further configured to: obtain the residual data of the (j+1)th block based on the extracted second block temporal features of the (j+1)th block and the residual data of the (j)th block; if the (j+1)th block is not the last block in the stack sorted from top to bottom, determine the residual data of the (j+1)th block as the input data of the (j+2)th block; if the (j+1)th block is the last block in the stack sorted from top to bottom, determine the residual data of the (j+1)th block as the stack residual data of the stack.

[0310] In this embodiment of the disclosure, the training module 124 is further configured to: obtain the training loss of the load prediction model based on the target load time series data and the sample load time series data; adjust the model parameters of the load prediction model based on the training loss; and return to continue training the adjusted load prediction model using the next sample load time series data and the next sample exogenous time series data until the training ends, thereby obtaining the target load prediction model.

[0311] The training apparatus for the load forecasting model proposed in this disclosure acquires the sample time-series features of sample load time-series data and the sample exogenous time-series features. Based on the sample time-series features and the sample exogenous features, load forecasting is performed within the sample time range to acquire second load time-series data within the sample time range. Based on the first and second load time-series data, the target load time-series data obtained by the load forecasting model within the sample time range is obtained. Further, based on the target load time-series data within the sample time range and the sample load time-series data input to the load forecasting model to be trained, the training loss of the load forecasting model to be trained is acquired. Based on the acquired training loss, the parameters of the load forecasting model are adjusted. After the model parameter adjustment is completed, the model training continues using the next sample load time-series data and the next sample exogenous time-series data until the training is completed and the trained target load forecasting model is obtained. In this disclosure, a load forecasting model is trained based on sample load time-series data and sample exogenous time-series data within a sample time range. This enables the load forecasting model to support the joint modeling of multiple variables and learn the correlation features between changes in exogenous data and load data. Consequently, the load forecasting model can obtain second load time-series data within the sample time range based on sample time-series features and sample exogenous features. Based on the first load time-series data and the second load time-series data, the final load forecasting model obtains the target load time-series data for load forecasting within the sample time range. This achieves load forecasting that combines features in the time-series dimension with features in the exogenous data dimension, optimizes the training method of the load forecasting model, improves the accuracy of the trained target load forecasting model for load forecasting, optimizes the applicability and practicality of the trained target load forecasting model, and optimizes the load forecasting method.

[0312] Corresponding to the load timing prediction methods proposed in the above embodiments, an embodiment of this disclosure also proposes a load timing prediction device. Since the load timing prediction device proposed in this disclosure corresponds to the load timing prediction methods proposed in the above embodiments, the implementation methods of the above load timing prediction methods are also applicable to the load timing prediction device proposed in this disclosure, and will not be described in detail in the following embodiments.

[0313] Figure 13 This is a schematic diagram of the structure of a load timing prediction device according to an embodiment of the present disclosure, as shown below. Figure 13 As shown, the load timing prediction device 1300 includes a third acquisition module 131 and a fourth prediction module 132, wherein:

[0314] The third acquisition module 131 is used to acquire historical load time series data to be predicted, as well as predictable exogenous time series data within the prediction time range.

[0315] The fourth prediction module 132 is used to input historical load time series data and predictable exogenous time series data into the trained target load prediction model, and obtain the target predicted load time series data within the prediction time range based on historical load time series data and predictable exogenous time series data from the output of the target load prediction model.

[0316] The target load prediction model is obtained based on the training device of the load prediction model proposed in the above embodiments.

[0317] The load time-series forecasting device disclosed herein acquires historical load time-series data to be forecasted, as well as predictable exogenous time-series data within the forecast time range, and inputs them into a trained target load forecasting model. The target forecast load time-series data within the forecast time range is obtained through the output of the target load forecasting model. In this disclosure, load forecasting is performed using a trained target load forecasting model, achieving load forecasting by combining exogenous time-series data and historical load time-series data, effectively improving the accuracy of load forecasting and optimizing the load forecasting method.

[0318] According to embodiments of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0319] Figure 14 A schematic block diagram of an example electronic device 1400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0320] like Figure 14As shown, device 1400 includes a computing unit 1401, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1402 or a computer program loaded from storage unit 1408 into random access memory (RAM) 1403. The RAM 1403 may also store various programs and data required for the operation of device 1400. The computing unit 1401, ROM 1402, and RAM 1403 are interconnected via bus 1404. Input / output (I / O) interface 1405 is also connected to bus 1404.

[0321] Multiple components in device 1400 are connected to I / O interface 1405, including: input unit 1406, such as keyboard, mouse, etc.; output unit 1406, such as various types of monitors, speakers, etc.; storage unit 1408, such as disk, optical disk, etc.; and communication unit 1409, such as network card, modem, wireless transceiver, etc. Communication unit 1409 allows device 1400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0322] The computing unit 1401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1401 performs the various methods and processes described above, such as load forecasting methods, load forecasting model training methods, and load timing forecasting methods. For example, in some embodiments, the load forecasting methods, load forecasting model training methods, and load timing forecasting methods can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 1408. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1400 via ROM 1402 and / or communication unit 1409. When the computer program is loaded into RAM 1403 and executed by computing unit 1401, one or more steps of the load forecasting method, the load forecasting model training method, and the load timing forecasting method described above can be performed. Alternatively, in other embodiments, computing unit 1401 can be configured to perform the load forecasting method, the load forecasting model training method, and the load timing forecasting method by any other suitable means (e.g., by means of firmware).

[0323] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0324] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0325] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0326] To facilitate interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to facilitate interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0327] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0328] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0329] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0330] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A training method for a load forecasting model, wherein, The method includes: Acquire time-series data of sample load and exogenous time-series data within the sample time range; The sample load time series data and the sample exogenous time series data are input into the load prediction model to be trained. The load prediction model obtains the first sample load change information of the sample load time series data under the influence of the sample exogenous time series data, and extracts the first sample feature of the first sample load change information. Based on the first sample characteristics, the first load time series data of the sample time range under the influence of the exogenous time series data of the sample is obtained; Based on the first load time series data and the sample load time series data, the load prediction model is adjusted, and the adjusted load prediction model is returned to continue training to obtain the trained target load prediction model. If the stack is the first stack of the load forecasting model, the sample load time series data and the sample exogenous time series data are used as the stack input data of the first stack; if the stack is the (i+1)th stack of the load forecasting model, the stack residual data of the i-th stack is obtained, and the stack residual data of the i-th stack and the sample exogenous data are used as the stack input data of the (i+1)th stack, where i is a positive integer greater than or equal to 1.

2. The method according to claim 1, wherein, The step of adjusting the load prediction model based on the first load time-series data and the sample load time-series data, and then returning to continue training the adjusted load prediction model to obtain a trained target load prediction model, includes: Based on the load forecasting model, sample time-series features of the sample load time-series data are obtained, wherein the sample time-series features include at least one of the sample time-series trend features and sample seasonality features of the sample load time-series data; Based on the time-series characteristics of the samples, the second load time-series data within the time range of the samples obtained from the intermediate layer of the load prediction model are acquired. Based on the first load time series data and the second load time series data, the target load time series data output by the load prediction model is obtained; Based on the target load time series data and the sample load time series data, the load prediction model is adjusted, and the adjusted load prediction model is returned to continue training to obtain a trained target load prediction model.

3. The method according to claim 2, wherein, The step of obtaining the second load time series data within the sample time range obtained from the intermediate layer of the load forecasting model based on the sample time series characteristics includes: Obtain the exogenous features of the samples corresponding to the exogenous time-series data of the samples; Under the combined influence of the sample time-series features and the sample exogenous features, obtain the second sample load change information of the sample load time-series data, and extract the second sample features corresponding to the second sample load change information; Based on the second sample characteristics, the second load time series data within the sample time range is obtained.

4. The method according to claim 3, wherein, The step of obtaining the target load time series data output by the load forecasting model based on the first load time series data and the second load time series data includes: Obtain the first load time series data output from at least one stack in a plurality of cascaded stacks of the load forecasting model, and the second load time series data output from the remaining stacks other than the at least one stack; The first load time series data and the second load time series data are integrated along the time series dimension to obtain the integrated target load time series data.

5. The method according to claim 4, wherein, The step of acquiring the first load time-series data output from at least one stack in a plurality of cascaded stacks of the load forecasting model, and the second load time-series data output from the remaining stacks other than the at least one stack, includes: Obtain the component load time series data output by each block in the multiple cascaded blocks of each stack in the load prediction model; Based on the component load timing data output by each block in each stack, obtain the first load timing data output by the at least one stack, and the second load timing data output by the remaining stacks other than the at least one stack.

6. The method according to claim 5, wherein, The process of acquiring the component load time series data includes: If the block is the first block in the stack, the stack input data is input into the first block to obtain the first component load timing data output by the first block; If the block is the (j+1)th block in the stack, obtain the residual data of the jth block, and input the residual data of the jth block and the sample exogenous time series data into the (j+1)th block to obtain the second component load time series data output by the (j+1)th block, where j is a positive integer greater than or equal to 1.

7. The method according to claim 6, wherein, The process of outputting the first component of the load time series data from the first block includes: The first block temporal features in the sample load time series data and the first block exogenous features in the sample exogenous time series data are extracted through the first block; From the output branch of the first block, obtain the first component load time series data within the sample time range obtained by the first block based on the time series features of the first block and the exogenous features of the first block.

8. The method according to claim 7, wherein, The process of outputting the second component load time-series data of the (j+1)th block includes: The second block temporal features are extracted from the residual data of the j-th block through the (j+1)-th block, and the second block exogenous features are extracted from the sample exogenous temporal data. From the output branch of the (j+1)th block, obtain the second component load time series data within the sample time range obtained by the (j+1)th block based on the time series features of the second block and the exogenous features of the second block.

9. The method according to claim 8, wherein, The method further includes: Based on the extracted temporal features of the second block from the (j+1)th block and the residual data of the jth block, obtain the residual data of the (j+1)th block; If the (j+1)th block is not the last block in the stack ordered from top to bottom, the residual data of the (j+1)th block is determined as the input data of the (j+2)th block; If the (j+1)th block is the last block of the stack sorted from top to bottom, then the residual data of the (j+1)th block is determined as the stack residual data of the stack.

10. The method according to any one of claims 1-9, wherein, The step of adjusting the load prediction model based on the target load time-series data and the sample load time-series data, and then returning to train the adjusted load prediction model to obtain a trained target load prediction model, includes: The training loss of the load prediction model is obtained based on the target load time series data and the sample load time series data. The model parameters of the load prediction model are adjusted according to the training loss, and the model is trained again using the next sample load time series data and the next sample exogenous time series data until the training ends, and the target load prediction model is obtained.

11. A load forecasting method, wherein, The method includes: Acquire historical load time series data, historical exogenous time series data, and predictable exogenous time series data within the forecast time range; Obtain first load change information of the historical load time series data under the influence of the historical exogenous time series data, and extract the first feature corresponding to the first load change information; Based on the first feature, obtain the first predicted load time series data within the predicted time range under the influence of the predictable exogenous time series data; Based on the first predicted load time series data, obtain the target predicted load time series data within the predicted time range; The method further includes: determining the target predicted load time series data based on the target load prediction model trained by the method of any one of claims 1-10.

12. The method according to claim 11, wherein, The step of obtaining the target predicted load time series data within the predicted time range based on the first predicted load time series data includes: Obtain the historical time-series features corresponding to the historical load time-series data, wherein the historical time-series features include at least one of historical time-series trend features and historical seasonality features; Based on the historical time series characteristics, obtain the second predicted load time series data within the predicted time range; The target predicted load time series data is obtained based on the first predicted load time series data and the second predicted load time series data.

13. The method according to claim 12, wherein, The step of obtaining the second predicted load time series data within the predicted time range based on the historical time series characteristics includes: Obtain the historical exogenous characteristics corresponding to historical exogenous time series data; Obtain the second load change information of the historical load time series data under the combined influence of the historical time series features and the historical exogenous features, and extract the second feature corresponding to the second load change information; Based on the second feature, the second predicted load time series data within the predicted time range is obtained.

14. The method according to claim 13, wherein, The step of obtaining the second predicted load time series data within the predicted time range based on the second feature includes: Obtain the predicted time series features corresponding to the predicted time range, wherein the predicted time series features include at least one of the predicted time series trend features and the predicted seasonality features corresponding to the predicted time range; From the predictable exogenous time series data, obtain the predictable exogenous features within the predicted time range; Based on the second feature, the second predicted load time series data within the predicted time range is obtained under the combined influence of the predicted time series feature and the predictable exogenous feature.

15. The method according to any one of claims 11-14, wherein, The process of acquiring predictable exogenous time series data within the predicted time range includes: Acquire time-series data of observed variables, predicted covariates, and static variables within the predicted time range; Based on the time series information within the predicted time range, the time series data of the observed variables, the predicted exogenous time series data, and the time series data of the static variables are integrated along the time series dimension to obtain the predictable exogenous time series data within the predicted time range.

16. The method according to any one of claims 12-14, wherein, The step of obtaining the target predicted load time series data within the predicted time range based on the first predicted load time series data and the second predicted load time series data includes: The first predicted load time series data and the second predicted load time series data are integrated according to the time series to obtain integrated load time series data, which is used as the target predicted load time series data.

17. A load time-series forecasting method, wherein, The method includes: Acquire historical load time series data to be predicted, as well as predictable exogenous time series data within the prediction time range; Input the historical load time series data and the predictable exogenous time series data into the trained target load prediction model, and obtain the target predicted load time series data within the prediction time range from the output of the target load prediction model, based on the historical load time series data and the predictable exogenous time series data. The target load prediction model is obtained based on the training method of the load prediction model according to any one of claims 1-10.

18. A load forecasting device, wherein, The device includes: The first acquisition module is used to acquire historical load time series data, historical exogenous time series data, and predictable exogenous time series data within the prediction time range; The first extraction module is used to obtain the first load change information of the historical load time series data under the influence of the historical exogenous time series data, and extract the first feature corresponding to the first load change information; The first prediction module is used to obtain, based on the first feature, the first predicted load time series data within the prediction time range under the influence of the predictable exogenous time series data; The second prediction module is used to obtain the target predicted load time series data within the prediction time range based on the first predicted load time series data; The apparatus further includes: determining the target predicted load time series data based on the target load prediction model trained by the method of any one of claims 1-10.

19. The apparatus according to claim 18, wherein, The second prediction module is further configured to: Obtain the historical time-series features corresponding to the historical load time-series data, wherein the historical time-series features include at least one of historical time-series trend features and historical seasonality features; Based on the historical time series characteristics, obtain the second predicted load time series data within the predicted time range; The target predicted load time series data is obtained based on the first predicted load time series data and the second predicted load time series data.

20. The apparatus according to claim 19, wherein, The second prediction module is further configured to: Obtain the historical exogenous characteristics corresponding to historical exogenous time series data; Obtain the second load change information of the historical load time series data under the combined influence of the historical time series features and the historical exogenous features, and extract the second feature corresponding to the second load change information; Based on the second feature, the second predicted load time series data within the predicted time range is obtained.

21. The apparatus according to claim 20, wherein, The second prediction module is further configured to: Obtain the predicted time series features corresponding to the predicted time range, wherein the predicted time series features include at least one of the predicted time series trend features and the predicted seasonality features corresponding to the predicted time range; From the predictable exogenous time series data, obtain the predictable exogenous features within the predicted time range; Based on the second feature, the second predicted load time series data within the predicted time range is obtained under the combined influence of the predicted time series feature and the predictable exogenous feature.

22. The apparatus according to any one of claims 18-21, wherein, The first acquisition module is further configured to: Acquire time-series data of observed variables, predicted covariates, and static variables within the predicted time range; Based on the time series information within the predicted time range, the time series data of the observed variables, the predicted exogenous time series data, and the time series data of the static variables are integrated along the time series dimension to obtain the predictable exogenous time series data within the predicted time range.

23. The apparatus according to any one of claims 18-21, wherein, The second prediction module is further configured to: The first predicted load time series data and the second predicted load time series data are integrated according to the time series to obtain integrated load time series data, which is used as the target predicted load time series data.

24. A training device for a load forecasting model, wherein, The device includes: The second acquisition module is used to acquire sample load time series data and sample exogenous time series data within the sample time range; The second extraction module is used to input the sample load time series data and the sample exogenous time series data into the load prediction model to be trained. The load prediction model obtains the first sample load change information of the sample load time series data under the influence of the sample exogenous time series data, and extracts the first sample feature of the first sample load change information. The third prediction module is used to obtain, based on the first sample characteristics, the first load time series data of the sample time range under the influence of the exogenous time series data of the sample; The training module is used to adjust the load prediction model based on the first load time series data and the sample load time series data, and return to continue training the adjusted load prediction model to obtain the trained target load prediction model. The stack extraction and prediction module is used to: if the stack is the first stack of the load prediction model, use the sample load time series data and the sample exogenous time series data as the stack input data of the first stack; if the stack is the (i+1)th stack of the load prediction model, obtain the stack residual data of the i-th stack, and use the stack residual data of the i-th stack and the sample exogenous data as the stack input data of the (i+1)th stack, where i is a positive integer greater than or equal to 1.

25. The apparatus according to claim 24, wherein, The training module is also used for: Based on the load forecasting model, sample time-series features of the sample load time-series data are obtained, wherein the sample time-series features include at least one of the sample time-series trend features and sample seasonality features of the sample load time-series data; Based on the time-series characteristics of the samples, the second load time-series data within the time range of the samples obtained from the intermediate layer of the load prediction model are acquired. Based on the first load time series data and the second load time series data, the target load time series data output by the load prediction model is obtained; Based on the target load time series data and the sample load time series data, the load prediction model is adjusted, and the adjusted load prediction model is returned to continue training to obtain a trained target load prediction model.

26. The apparatus according to claim 25, wherein, The training module is also used for: Obtain the exogenous features of the samples corresponding to the exogenous time-series data of the samples; Under the combined influence of the sample time-series features and the sample exogenous features, obtain the second sample load change information of the sample load time-series data, and extract the second sample features corresponding to the second sample load change information; Based on the second sample characteristics, the second load time series data within the sample time range is obtained.

27. The apparatus according to claim 26, wherein, The training module is also used for: Obtain the first load time series data output from at least one stack in a plurality of cascaded stacks of the load forecasting model, and the second load time series data output from the remaining stacks other than the at least one stack; The first load time series data and the second load time series data are integrated along the time series dimension to obtain the integrated target load time series data.

28. The apparatus according to claim 27, wherein, The stack extraction and prediction module is also used for: Obtain the component load time series data output by each block in the multiple cascaded blocks of each stack in the load prediction model; Based on the component load timing data output by each block in each stack, obtain the first load timing data output by the at least one stack, and the second load timing data output by the remaining stacks other than the at least one stack.

29. The apparatus according to claim 28, wherein, The stack extraction and prediction module is also used for: If the block is the first block in the stack, the stack input data is input into the first block to obtain the first component load timing data output by the first block; If the block is the (j+1)th block in the stack, obtain the residual data of the jth block, and input the residual data of the jth block and the sample exogenous time series data into the (j+1)th block to obtain the second component load time series data output by the (j+1)th block, where j is a positive integer greater than or equal to 1.

30. The apparatus according to claim 29, wherein, The stack extraction and prediction module is also used for: The first block temporal features in the sample load time series data and the first block exogenous features in the sample exogenous time series data are extracted through the first block; From the output branch of the first block, obtain the first component load time series data within the sample time range obtained by the first block based on the time series features of the first block and the exogenous features of the first block.

31. The apparatus according to claim 30, wherein, The stack extraction and prediction module is also used for: The second block temporal features are extracted from the residual data of the j-th block through the (j+1)-th block, and the second block exogenous features are extracted from the sample exogenous temporal data. From the output branch of the (j+1)th block, obtain the second component load time series data within the sample time range obtained by the (j+1)th block based on the time series features of the second block and the exogenous features of the second block.

32. The apparatus according to claim 31, wherein, The stack extraction and prediction module is also used for: Based on the extracted temporal features of the second block from the (j+1)th block and the residual data of the jth block, obtain the residual data of the (j+1)th block; If the (j+1)th block is not the last block in the stack ordered from top to bottom, the residual data of the (j+1)th block is determined as the input data of the (j+2)th block; If the (j+1)th block is the last block of the stack sorted from top to bottom, then the residual data of the (j+1)th block is determined as the stack residual data of the stack.

33. The apparatus according to any one of claims 25-32, wherein, The training module is also used for: The training loss of the load prediction model is obtained based on the target load time series data and the sample load time series data. The model parameters of the load prediction model are adjusted according to the training loss, and the model is trained again using the next sample load time series data and the next sample exogenous time series data until the training ends, and the target load prediction model is obtained.

34. A load timing prediction device, wherein, The device includes: The third acquisition module is used to acquire historical load time series data to be predicted, as well as predictable exogenous time series data within the prediction time range; The fourth prediction module is used to input the historical load time series data and the predictable exogenous time series data into the trained target load prediction model, and obtain the target predicted load time series data within the prediction time range based on the historical load time series data and the predictable exogenous time series data from the output of the target load prediction model. The target load prediction model is obtained based on the training device of the load prediction model according to any one of claims 24-33.

35. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-17.

36. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-17.

37. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-17.

Images