CN-122000864-A - Cross-scene power load prediction method, device, system and medium under large model

CN122000864ACN 122000864 ACN122000864 ACN 122000864ACN-122000864-A

Abstract

The embodiment of the application provides a method, a device, a system and a medium for predicting a cross-scene power load under a large model, and relates to the technical field of power load prediction. The method comprises the steps of generating a cross-scene power load association feature set according to specific power consumption type power load data of a source scene and a target scene and corresponding power consumption factor data, adjusting weight parameters of a pre-trained power load prediction big model according to the cross-scene power load association feature set, converting the prediction parameters of the source scene into load prediction adaptation parameters for adapting to the target scene through migration learning, and carrying out real-time load feature optimization and prediction calculation on the target scene by utilizing the big model and the load prediction adaptation parameters after the weight parameters are adjusted to generate a target scene specific power consumption type power load prediction result. The method improves the accuracy and adaptability of the cross-scene power load prediction.

Inventors

LIU XIN
Sun Mengqian
LIU ZIYUAN
LIU DONGLAN
WANG RUI
ZHANG HAO
MA LEI
ZHAO FUHUI
XU SHANJIE
ZHANG FANGZHE
YU HAO

Assignees

国网山东省电力公司电力科学研究院

Dates

Publication Date: 20260508
Application Date: 20251215

Claims (16)

1. A method for cross-scene power load prediction under a large model, comprising: Generating a cross-scene power load association feature set according to specific power utilization type power load data of a source scene and a target scene and corresponding power utilization factor data; According to the cross-scene power load association feature set, adjusting the weight parameters of a pre-trained power load prediction large model, and converting the prediction parameters of a source scene into load prediction adaptation parameters for adapting to a target scene through transfer learning; And carrying out real-time load characteristic optimization and prediction calculation on the target scene by utilizing the large model and the load prediction adaptive parameters after the weight parameters are adjusted, and generating a power load prediction result of the specific power utilization type of the target scene.
2. The method of claim 1, wherein generating a cross-scenario power load association feature set from power load data of a particular power usage type of a source scenario and a target scenario and corresponding power factor data comprises: extracting a load value sequence corresponding to the same time period identifier from the source scene specific electricity consumption type power load data set, calculating the variation amplitude of the load value under the same time period identifier of different dates, and generating source scene power load time period association characteristics; Extracting a load value sequence corresponding to the same time period identifier from the target scene specific electricity consumption type power load data set, calculating the variation amplitude of the load value under the same time period identifier of different dates, and generating a target scene power load time period association characteristic; And the source scene power load time period association characteristic and the target scene power load time period association characteristic are in one-to-one correspondence according to the time period identification, similarity parameters of the source scene power load time period association characteristic and the target scene power load time period association characteristic under the corresponding time period identification are calculated, and the cross-scene load time period association characteristic is generated.
3. The method of claim 2, wherein extracting a sequence of load values corresponding to the same time period identifier from the source scenario-specific electricity type power load data set, calculating a variation amplitude of the load values under the same time period identifier of different dates, and generating source scenario power load time period association features, comprises: Screening all load data with the same time period identification from the source scene specific electricity consumption type power load data set to form a source scene load value sequence, and calculating load change amplitudes of all adjacent dates according to load values of two adjacent dates in the source scene load value sequence to form a load change amplitude sequence; and acquiring an ascending trend characteristic value, a descending trend characteristic value and a smooth trend characteristic value according to the load change amplitude sequence, and integrating and forming the source scene power load period association characteristic.
4. The method of claim 2, wherein generating a cross-scenario power load association feature set from power load data of a particular power usage type of a source scenario and a target scenario and corresponding power factor data comprises: Extracting association relation between meteorological data and industrial production shift data from the dynamic factor data for the source scene, calculating load change rates corresponding to industrial production shifts under different meteorological data, and generating dynamic factor association characteristics for the source scene; extracting association relation between meteorological data and resident life information data from the electric response factor data for the target scene, calculating load change rates corresponding to resident life information under different meteorological data, and generating electric response factor association characteristics for the target scene; And the source-use dynamic response factor association characteristic and the target-use dynamic response factor association characteristic are in one-to-one correspondence according to the meteorological data type, complementary parameters of the source-use dynamic response factor association characteristic and the target-use dynamic response factor association characteristic under the corresponding meteorological data type are calculated, and the cross-scene load influence factor association characteristic is generated.
5. The method of claim 4, wherein adjusting weight parameters of a pre-trained power load prediction large model according to the cross-scene power load correlation feature set and converting the predicted parameters of the source scene into load prediction adaptation parameters for adapting the target scene by transfer learning comprises: Inputting the cross-scene load period association features into the large model, carrying out hierarchical analysis on source scene period features, target scene period features and similarity parameters in the cross-scene load period association features through the large model, and identifying significant period feature dimensions aiming at load prediction precision; And calculating the contribution ratio of the feature dimension of the significant period in the source scene and the target scene, and adjusting the feature extraction weight of the corresponding dimension in the large model according to the contribution ratio.
6. The method of claim 5, wherein adjusting weight parameters of a pre-trained power load prediction large model according to the cross-scene power load correlation feature set and converting the predicted parameters of the source scene to load prediction adaptation parameters for adapting the target scene by transfer learning comprises: Inputting the cross-scene load influence factor association features into the large model, carrying out hierarchical analysis on the source scene influence factor features, the target scene influence factor features and the complementary parameters in the cross-scene load influence factor association features through the large model, and identifying significant influence factor feature dimensions aiming at load prediction precision; And calculating the complementary coefficient of the characteristic dimension of the significant influence factor in the source scene and the target scene, and adjusting the characteristic extraction weight of the corresponding dimension in the large model according to the complementary coefficient.
7. The method of claim 6, wherein adjusting weight parameters of a pre-trained power load prediction large model according to the cross-scene power load correlation feature set and converting the predicted parameters of the source scene to load prediction adaptation parameters for adapting the target scene by transfer learning comprises: According to the time period characteristics and the influence factor characteristics of the source scene and the target scene after the weight adjustment, and according to scene difference parameters in the cross-scene power load association characteristic set, adjusting the fusion proportion of the time period characteristics and the influence factor characteristics, and generating a cross-scene fusion load characteristic; inputting the cross-scene fusion load characteristics into the large model, verifying the load prediction error of the cross-scene fusion load characteristics, and if the error exceeds a preset range, readjusting the characteristic extraction weight of the corresponding dimension until the prediction error accords with the preset range.
8. The method of claim 7, wherein adjusting the fusion ratio of the time period feature and the influence factor feature according to the time period feature and the influence factor feature of the source scene and the target scene after the weight adjustment and according to the scene difference parameter in the cross-scene power load association feature set, and generating a cross-scene fusion load feature comprises: Acquiring output time period characteristics and influence factor characteristics from the large model after weight adjustment, and generating a cross-scene fusion load characteristic vector according to the time period characteristics and the influence factor characteristics; and integrating all time period identifications and the cross-scene fusion load characteristic vectors corresponding to the type of the dynamic factor to form the cross-scene fusion load characteristic.
9. The method of claim 7, wherein adjusting weight parameters of a pre-trained power load prediction large model according to the cross-scene power load correlation feature set and converting the predicted parameters of the source scene to load prediction adaptation parameters for adapting the target scene by transfer learning comprises: Extracting a source scene power load prediction parameter from the large model, performing scene difference adaptation processing on time period feature extraction weights in the source scene power load prediction parameter, calculating adjustment proportion of the time period feature extraction weights based on load peak value differences in scene difference features, if the load peak value of a certain time period of a target scene is higher than a corresponding time period of the source scene, improving the time period feature extraction weights corresponding to the corresponding time period, otherwise, reducing the time period feature extraction weights corresponding to the time period, and obtaining the adapted time period feature extraction weights.
10. The method of claim 9, wherein adjusting weight parameters of a pre-trained power load prediction large model according to the cross-scene power load correlation feature set and converting the predicted parameters of the source scene into load prediction adaptation parameters for adapting the target scene by transfer learning comprises: And performing scene difference adaptation processing on the influence factor feature extraction weight in the source scene power load prediction parameters, calculating an adjustment coefficient of the influence factor feature extraction weight based on the load change rate difference in the scene difference feature, if the load change rate of certain meteorological data of the target scene is higher than the change rate of the meteorological data corresponding to the source scene, improving the influence factor feature extraction weight corresponding to the meteorological data, otherwise, reducing the influence factor feature extraction weight corresponding to the meteorological data, and obtaining the adapted influence factor feature extraction weight.
11. The method of claim 10, wherein adjusting weight parameters of a pre-trained power load prediction large model according to the cross-scene power load correlation feature set and converting the predicted parameters of the source scene to load prediction adaptation parameters for adapting the target scene by transfer learning comprises: Performing scene difference adaptation processing on the characteristic fusion coefficient in the power load prediction parameters of the source scene, calculating a correction value of the characteristic fusion coefficient based on the load fluctuation period difference in the scene difference characteristic, if the load fluctuation period of the target scene is longer than that of the source scene, increasing the duty ratio of the time period characteristic in fusion, correcting the characteristic fusion coefficient, otherwise, increasing the duty ratio of the influence factor characteristic in fusion, and correcting the characteristic fusion coefficient to obtain the adapted characteristic fusion coefficient; Performing scene difference adaptation processing on the prediction calculation parameters in the source scene power load prediction parameters, calculating compensation values of the prediction calculation parameters based on the comprehensive influence of load peak value differences, load change rate differences and load fluctuation period differences in scene difference characteristics, and overlapping the compensation values on the source scene prediction calculation parameters to obtain adapted prediction calculation parameters; Integrating the time period feature extraction weight, the influence factor feature extraction weight, the feature fusion coefficient and the prediction calculation parameter after the adaptation to form a load prediction adaptation parameter adapting to the specific power utilization type of the target scene; substituting the load prediction adaptive parameters into historical load data samples of the specific electricity utilization type of the target scene, executing prediction verification calculation, and if the deviation between a prediction result and the historical actual load data exceeds a preset threshold value, returning to recalculate the parameter adjustment proportion, the adjustment coefficient, the correction value and the compensation value until the deviation accords with the preset threshold value.
12. The method of claim 11, wherein generating a target scenario-specific electricity type power load prediction result by performing real-time load feature optimization and prediction calculation for the target scenario using the large model and the load prediction adaptation parameters after the weight parameters are adjusted, comprises: Acquiring real-time specific electricity utilization type power load data and corresponding real-time electric utilization factor data of a target scene, acquiring real-time period association features and real-time influence factor association features through the large model, and further generating real-time fusion load features; according to the real-time fusion load characteristics, carrying out load numerical prediction through the large model to form a specific electricity consumption type power load prediction sequence of a target scene; and extracting a load peak value, a load valley value and a load average value of each prediction period from the target scene specific electricity type power load prediction sequence, and integrating the target scene specific electricity type power load prediction sequence and the prediction characteristic parameter to generate a target scene specific electricity type power load prediction result.
13. The method of claim 11, wherein generating a target scenario-specific electricity type power load prediction result by performing real-time load feature optimization and prediction calculation for the target scenario using the large model and the load prediction adaptation parameters after the weight parameters are adjusted, comprises: And generating a load adjustment instruction for target scene power scheduling based on the target scene specific electricity type power load prediction result, wherein the content of the load adjustment instruction corresponds to a load period change rule in the target scene specific electricity type power load prediction result.
14. A cross-scene power load prediction device under a large model, comprising a processor and a memory storing program instructions, wherein the processor is configured to perform the cross-scene power load prediction method under a large model as claimed in any one of claims 1 to 13 when the program instructions are run.
15. A system, comprising: system body, and The large model cross-scenario power load prediction apparatus according to claim 14, mounted to the system body.
16. A computer readable storage medium storing program instructions which, when executed, are to cause a computer to perform the method of cross-scene power load prediction under a large model according to any one of claims 1 to 13.

Description

Cross-scene power load prediction method, device, system and medium under large model Technical Field The application relates to the technical field of power load prediction, in particular to a method, a device, a system and a storage medium for predicting a cross-scene power load under a large model. Background In a power system, accurate power load prediction is critical to power scheduling, resource allocation, and ensuring stable operation of the power system. Traditional power load prediction methods are mainly based on historical load data in a single scene for analysis and modeling. For example, a time series analysis method such as an autoregressive moving average model (ARMA) or the like is used to predict the future power load based only on the power load values of a certain region over a period of time. However, the above-described single scene prediction method has significant limitations. The power load in different scenarios is affected by a number of factors that behave differently in different scenarios. For example, the power load in an industrial scene is mainly affected by factors such as shifts in industrial production and operating states of equipment, while the power load in a resident scene is closely related to resident life information, weather conditions, and the like. The prediction method of a single scene cannot sufficiently consider these differences, resulting in low prediction accuracy in different scenes. In addition, when the mature prediction model in a certain scene is required to be applied to other scenes, the traditional method cannot be used by direct migration due to the difference between the scenes, and a large amount of data collection and model training are required to be carried out again, so that a large amount of time and resources are consumed, and the requirements of different scenes are difficult to adapt quickly. Therefore, the existing power load prediction method has obvious defects in the aspect of cross-scene application, and cannot meet the increasingly complex and diversified requirements of a power system. Disclosure of Invention The embodiment of the application provides a method, a device, a system and a storage medium for predicting cross-scene power load under a large model. In a first aspect of the embodiment of the present application, a method for predicting a cross-scene power load under a large model is provided, including: Generating a cross-scene power load association feature set according to specific power utilization type power load data of a source scene and a target scene and corresponding power utilization factor data; According to the cross-scene power load association feature set, adjusting the weight parameters of a pre-trained power load prediction large model, and converting the prediction parameters of a source scene into load prediction adaptation parameters for adapting to a target scene through transfer learning; and carrying out real-time load characteristic optimization and prediction calculation aiming at the target scene by utilizing the large model with the weight parameters adjusted and the load prediction adaptive parameters, and generating a power load prediction result of the specific power utilization type of the target scene. In an alternative embodiment of the present application, generating a cross-scenario power load association feature set according to specific power usage type power load data of a source scenario and a target scenario and corresponding application factor data includes: extracting a load value sequence corresponding to the same time period identifier from a source scene specific electricity consumption type power load data set, calculating the variation amplitude of load values under the same time period identifier of different dates, and generating source scene power load time period association characteristics; Extracting a load value sequence corresponding to the same time period identifier from a target scene specific electricity consumption type power load data set, calculating the variation amplitude of load values under the same time period identifier of different dates, and generating a target scene power load time period association characteristic; And (3) corresponding the source scene power load time period association features and the target scene power load time period association features one by one according to the time period identification, calculating similarity parameters of the source scene power load time period association features and the target scene power load time period association features under the corresponding time period identification, and generating cross-scene load time period association features. In an optional embodiment of the present application, extracting a load value sequence corresponding to the same period identifier from a source scene specific electricity type power load data set, calculating variation amplitudes of load values under the same per