Electrical load prediction method and system based on multi-source data fusion

Abstract

The invention discloses an electrical load prediction method and system based on multi-source data fusion, and relates to the technical field of electrical load prediction. The method comprises the steps that a power utilization planning request is sent to a back-end API through an HTTP protocol, the API interface receives planning request data sent by a front end, and a load predictor is loaded; according to a load predictor, dual-mode planning is carried out on the planning request data, a load prediction curve is determined, and the load predictor carries out forward and reverse training construction based on an element layer, a semantic layer and an intention layer; and on a front-end interface, the load prediction curve is visualized through an API interface. The technical problems that in the prior art, due to the fact that a power utilization load prediction model is insufficient in coupling of multi-source data and user planning intentions, prediction results are difficult to meet actual power utilization planning requirements, and usability is poor are solved. The technical effect of improving the availability of the electrical load prediction result is achieved.

Description

Technical Field

[0001] This invention relates to the field of electricity load forecasting technology, and specifically to an electricity load forecasting method and system based on multi-source data fusion. Background Technology

[0002] With the increasing demands for refined management of new power systems and electricity consumption, load forecasting has become a crucial foundational technology in electricity planning, operation scheduling, and energy efficiency management. Existing load forecasting methods primarily rely on historical electricity consumption data, meteorological data, or equipment operation data to construct forecasting models for future load trends. However, in practical applications, load forecasting often needs to be integrated with specific electricity planning scenarios. Users initiate planning requests based on business objectives, time-period requirements, or policy constraints, and the forecast results must directly support subsequent planning decisions and business applications. Furthermore, most existing load forecasting models focus on modeling single or limited data characteristics, lacking a unified fusion mechanism for data from different sources. They also typically fail to effectively depict the implicit planning intentions of users in their planning requests, leading to discrepancies between the load results generated by the forecasting models and actual planning needs, resulting in low usability and business adaptability. Additionally, existing load forecasting processes are mostly based on offline calculations or single forecasting modes, making it difficult to flexibly balance forecasting efficiency and accuracy, and failing to simultaneously meet the application requirements of rapid planning and refined forecasting. In addition, existing electricity load forecasting systems are often tightly coupled with front-end business applications at the system architecture level, resulting in insufficient ability to call, display, and interact with forecast results, which is not conducive to the reuse and visualization of forecast results in electricity planning systems. Summary of the Invention

[0003] This application provides a method and system for predicting electricity load based on multi-source data fusion, which solves the technical problem in the prior art where the electricity load prediction model does not sufficiently couple multi-source data with user planning intentions, resulting in prediction results that are difficult to meet actual electricity planning needs and have poor usability.

[0004] The first aspect of this application provides a method for predicting electricity load based on multi-source data fusion, the method comprising: A power planning request is sent to the backend API via the HTTP protocol. The API interface receives the planning request data sent by the frontend and loads the load forecaster. Based on the load forecaster, dual-mode planning is performed on the planning request data to determine the load forecast curve. The load forecaster is built based on the primitive layer, semantic layer, and intent layer through forward and reverse training. The dual modes include coarse planning and precise planning. The load forecast curve is visualized on the frontend interface through the API interface.

[0005] A second aspect of this application provides an electricity load forecasting system based on multi-source data fusion, the system comprising: Data sending module: Sends electricity planning requests to the backend API via HTTP protocol. The API interface receives the planning request data sent by the frontend and loads the load forecaster. Load forecasting module: Performs dual-mode planning on the planning request data based on the load forecaster to determine the load forecast curve. The load forecaster is built based on the primitive layer, semantic layer, and intent layer through forward and reverse training. The dual modes include coarse planning and precise planning. Data visualization module: Visualizes the load forecast curve on the frontend interface through the API interface.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, a power planning request is sent to the backend API via HTTP. The API receives the planning request data from the frontend and loads the load forecaster. Then, based on the load forecaster, dual-mode planning is performed on the planning request data to determine the load forecast curve. The load forecaster is built through forward and backward training based on a primitive layer, a semantic layer, and an intent layer. The dual modes include coarse planning and precise planning. Finally, the load forecast curve is visualized on the frontend interface via the API. This solves the technical problem in existing power load forecasting models where insufficient coupling between multi-source data and user planning intent leads to poor forecast results that fail to meet actual power planning needs and have poor usability. By constructing a load forecaster and performing dual-mode load forecasting on planning requests, the usability of power load forecasting results is improved. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic diagram of the electricity load forecasting method based on multi-source data fusion provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of an electricity load forecasting system based on multi-source data fusion provided in an embodiment of this application.

[0009] Figure labeling: Data transmission module 11, load forecasting module 12, data visualization module 13. Detailed Implementation

[0010] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0011] Example 1, as Figure 1 As shown, this application provides a method for predicting electricity load based on multi-source data fusion, wherein the method includes: The power planning request is sent to the backend API via the HTTP protocol. The API interface receives the planning request data sent by the frontend and loads the load forecaster.

[0012] In this embodiment, when sending a power planning request to the backend API via the HTTP protocol, the frontend encapsulates the planning request data describing the power planning needs into a standardized request message and sends it to the backend server via HTTP communication. The planning request data includes at least forecast period information, power consumption scenario identifiers, load constraint parameters, and user planning preference parameters. Upon receiving the planning request data, the pre-defined API interface in the backend server performs validity verification and parameter parsing on the request message, extracting key input information for load forecasting. After parameter parsing, the API interface triggers a load forecasting processing flow based on the request type, retrieves a load forecaster instance matching the power consumption scenario from the forecasting model management module, and loads the planning request data into the input buffer of the load forecaster, completing the loading and initialization configuration of the load forecaster. This provides a computing environment and model support for subsequent load forecasting calculations based on the planning request data.

[0013] Furthermore, the construction process of the load forecaster before loading it includes: Historical load curves are retrieved, and load elements, the smallest behavioral units, are deconstructed and determined. A causal inference chain based on external strong driving factors and internal load element activation is reconstructed for each load element to determine a first causal view. Association rules are then reconstructed for each load element to determine a second association view, where the association rules are the symbiotic, mutually exclusive, and sequential associations between load elements in terms of time sequence and spatial location. A generative power spectrum is defined using the first causal view and the second association view.

[0014] Specifically, historical load curve data is retrieved. This historical load curve includes at least a power value sequence recorded in chronological order and corresponding time identifiers. The historical load curve is then segmented and deconstructed to break down the continuous load change process into the smallest behavioral units that can characterize a single electricity consumption behavior or load state change, thus determining the load element. After obtaining the load element, the influence of external strong driving factors on load changes is introduced. These external strong driving factors include at least meteorological conditions, time attributes, electricity price information, or production plans. By analyzing the response relationship between changes in external strong driving factors and the activation state of the load element, a causal reasoning chain from external strong driving factors to the activation of internal load elements is reconstructed, forming a load element... A first causal view is constructed to describe the causal relationship of load generation. Simultaneously, statistical analysis is performed on the order of occurrence, spatial affiliation, and combination patterns of the load elements in historical load curves. This uncovers symbiotic, mutually exclusive, and sequential relationships among the load elements in terms of temporal and spatial relationships. Based on these relationships, association rules are constructed between the load elements to form a second association view. Furthermore, the first causal view and the second association view are modeled in a unified manner to define a generative power spectrum containing load element nodes and their causal and association relationships. This spectrum is used to characterize the generation mechanism and evolutionary constraints of electricity load under multi-source driving conditions, providing structured prior constraints for the training and prediction of subsequent load forecasters.

[0015] Based on the load forecaster, dual-mode planning is performed on the planning request data to determine the load forecast curve. The load forecaster is constructed by forward and reverse training based on the primitive layer, semantic layer and intent layer. The dual-mode includes coarse planning and precise planning.

[0016] Upon receiving the planning request data, the system selects the corresponding prediction mode and performs load planning processing based on the prediction period, electricity consumption scenario type, and constraints indicated in the planning request. In the coarse planning mode, the planning request data and real-time multi-source environmental data are mapped to the primitive layer and semantic layer. Using the primitive assembly rules and semantic generation paths obtained through forward training, load change trends are rapidly generated under the constraints of a generative power graph, resulting in a coarse load prediction curve for preliminary decision-making. In the precise planning mode, the intent layer inference results are further introduced to refine the coarse load prediction curve. The load intent reverse generation path formed through reverse training corrects the load action semantics, and the corrected semantic trajectory is instantiated into a high-time-resolution power time series, thereby generating a precise load prediction curve that meets the planning constraints. This dual-mode planning allows the load prediction process to balance prediction response speed and prediction accuracy at different planning stages, ultimately outputting a load prediction curve.

[0017] Furthermore, a multi-layer coding approach based on primitive layer, semantic layer and intent layer is adopted, and a load predictor is built by performing bidirectional training; wherein, the primitive layer is built based on the generative power map, the semantic layer is built based on load action phrases under primitive sequence combination, and the intent layer is built by power load intent reasoning based on load action phrases.

[0018] Preferably, the load forecaster is constructed using a multi-layer coding approach based on a primitive layer, a semantic layer, and an intent layer, and forms a model structure usable for load forecasting through bidirectional training. Specifically, the primitive layer is constructed based on a generative power graph, using load primitives in the generative power graph as the smallest coding unit. It combines the causal and correlational relationships between load primitives to structurally encode historical load behavior, forming a primitive representation that characterizes the basic patterns of load change. The semantic layer performs combined modeling based on the primitive sequence output from the primitive layer, combining multiple load primitives with temporal order and functional association into load action phrases describing segments of electricity consumption behavior. These load action phrases are then semantically encoded to characterize the behavioral meaning of load changes under different electricity consumption scenarios. The intent layer uses these load action phrases as inference input, analyzes the combination patterns between load action phrases and their correspondence with planning constraints, performs electricity load intent inference, and generates an intent representation that characterizes user electricity planning objectives and load adjustment tendencies. During the training process of the load forecaster, forward and reverse training are performed on the above multi-layer coding structure to optimize the mapping relationship between the primitive layer, semantic layer and intent layer, thereby constructing a load forecaster that can simultaneously characterize the load generation mechanism and planning intent constraints.

[0019] Furthermore, this is achieved by performing bidirectional training, including: Forward training based on the primitive layer, semantic layer, and intent layer is performed to generate a first prediction path, wherein the forward training is based on the primitive assembly splicing generation and semantic action intent reasoning based on the generative power spectrum; inverse training based on the primitive layer, semantic layer, and intent layer is performed to generate a second prediction path, wherein the inverse training is based on the load intent reverse generation under semantic trajectory probability, load action phrase decoding, and power time series instantiation.

[0020] A load predictor is constructed by performing bidirectional training, which includes two training processes: forward training and backward training. In the forward training process, based on a generative power graph, starting from the primitive layer, primitive assembly is performed according to the causal and correlational relationships between load primitives. The load primitives obtained from the decomposition of historical load curves are sequentially assembled into primitive sequences, and corresponding load action phrases are generated based on these primitive sequences. After generating the load action phrases, semantic action intent reasoning is further performed in the semantic and intent layers to learn the mapping relationship from the primitive sequence to the load action semantics and then to the power load intent. This forms the first prediction path, which deduces from load primitives layer by layer to the load intent, and is used to characterize the forward evolution process of load generation. During the reverse training process, based on the semantic trajectory distribution formed during the forward training process, semantic trajectory probability constraints are introduced into the intent layer. Starting from the power load intent, reverse generation reasoning of the load intent is performed, decoding the corresponding load action phrases and their primitive combinations layer by layer. The decoded load action phrases are further instantiated into power time series with time identifiers and power values, thus forming a second prediction path that reversely reconstructs the specific power change from the load intent. Through the coordinated execution of the above forward and reverse training, the load forecaster can simultaneously learn the reverse mapping relationship between the load generation process and the load planning constraints, improving the stability and usability of the load forecast results under different planning scenarios.

[0021] Furthermore, performing dual-mode planning on the planning request data to determine the load forecast curve includes: The system acquires planning request data, which is a load curve generation for a preset time period. The load curve generation is modeled as a sequence generation task based on graph constraints. Multi-source data is collected in real time to generate an external environment vector. Based on the planning request data and the external environment vector, short-term prediction is performed in the loaded load forecaster to generate a load forecast curve.

[0022] Preferably, the system receives and parses electricity planning requests from the front end to obtain planning request data. This data describes the load curve generation requirements within a preset forecast period. The load curve generation process is modeled as a sequence generation task constrained by a generative power graph, i.e., generating a load change sequence that conforms to electricity consumption patterns under the constraints of the evolutionary relationship between load primitives, load action semantics, and load intent. Simultaneously with obtaining the planning request data, multi-source data related to electricity load changes are collected in real time, including meteorological information, time attributes, electricity price information, or electricity scenario status information. This multi-source data undergoes unified dimension processing and feature encoding to form an external environment vector characterizing the current external operating conditions. After constructing the external environment vector, the planning request data and the external environment vector are input into the loaded load predictor to trigger a short-term prediction process. By calling the prediction path obtained through bidirectional training at the primitive layer, semantic layer, and intent layer, a corresponding load change sequence is generated under the constraints of the generative power graph. This load change sequence is then instantiated as a power value that changes over time, thereby outputting the load prediction curve.

[0023] Furthermore, short-term forecasting in the loaded load forecaster includes: By fusing external environment vectors and planning request data, and based on the first prediction path, the first candidate load map is determined by splicing primitive sequences in the generative power map, and the first load semantic trajectory is recursively deduced; based on the second prediction path, the second load semantic trajectory is determined by decoding, instantiated into the second power time series, and the coarse planning result is determined.

[0024] After receiving the planning request data, the system synchronously collects multi-source external environment data corresponding to the prediction period, and performs normalization and vectorization processing on the external environment data to generate an external environment vector. The external environment vector is then fused and encoded with the planning request data and input into the load forecaster to drive the first prediction path and the second prediction path to perform prediction calculations.

[0025] In the first prediction path, based on the forward generation capability of the primitive layer, under the causal and correlation constraints of the generative power spectrum, the load primitives are serialized and spliced ​​to generate candidate load primitive combinations that meet the planning request constraints. The candidate load primitive combinations are then used as a set of nodes to construct the first candidate load spectrum. On this basis, the primitive combinations in the first candidate load spectrum are semantically mapped and temporally recursively analyzed through the semantic layer to generate the first load semantic trajectory that describes the evolution of the load over time.

[0026] In the second prediction path, based on the inverse training capability of the load forecaster, the second load semantic trajectory is obtained by inverse decoding from the load intent or load semantic layer, using the fused encoded planning request data and the external environment vector as conditions. According to the power mapping relationship corresponding to each semantic state in the second load semantic trajectory, the second load semantic trajectory is instantiated into a power time series to generate a second power time series. The second power time series is used as the prediction output for the rapid response to the planning request to determine the coarse planning result, which is used to characterize the overall change trend of the electricity load within the prediction period.

[0027] Furthermore, after determining the rough planning results, this includes: Based on the rough planning results, the load intention periods and transferable loads with high electricity price elasticity are determined according to the load semantic trajectory. For the load intention periods and transferable loads, cross-time period optimization and redistribution are performed to determine the optimal electricity consumption plan and generate the load forecast curve. The cross-time period is divided into peak, flat, valley and deep valley as the basis for redistribution.

[0028] After determining the rough planning results, the adjustability characteristics of the electricity load are analyzed and processed based on the load semantic trajectory corresponding to the rough planning results. Specifically, according to the semantic features representing load change behavior in the load semantic trajectory, load intention periods that are sensitive to electricity price changes and have time adjustment space are identified, and transferable loads that can be time-shifted without affecting the electricity consumption function constraints are determined within the load intention periods. After obtaining the load intention periods and transferable loads, cross-time period optimization and redistribution processing is performed on the transferable loads with the optimization objective of reducing electricity costs or smoothing the load curve. The cross-time period optimization uses the time-of-use electricity price intervals of the power system's preset peak, mid-peak, normal, valley, and deep valley periods as the redistribution basis. By transferring the transferable loads from high-price periods to low-price periods, a new electricity load distribution that satisfies the load balance constraints and planning constraints is generated. After completing the cross-time period optimization and redistribution, a corresponding load prediction curve is generated based on the optimized electricity load distribution, which is output as the optimal electricity consumption planning result.

[0029] The load forecast curve can be visualized through the API interface on the front-end interface.

[0030] When visualizing the load forecast curve via API in the front-end interface, the back-end performs structured processing on the generated load forecast curve according to a preset data format. The load forecast curve includes at least the power values ​​and time markers corresponding to each time point within the forecast period, and optionally associates with corresponding optimization decision information. After structuring, the load forecast curve is returned to the front-end interface via API. Upon receiving the load forecast curve data, the front-end parses the data according to the time and power dimensions and displays it as a curve, segmented bar chart, or other visual graphics, thus intuitively presenting the changing trend of electricity load within the forecast period. Simultaneously, the front-end interface can support users in comparing or interactively analyzing different forecast results based on the load forecast curve, assisting users in making electricity planning decisions.

[0031] Furthermore, the load forecast curve is visualized on the front-end interface via an API interface, including: An optimization decision log is generated, in which the load transfer status and electricity cost comparison during different time periods are used as generation elements; the load forecast curve is structured and stored in a temporary database along with the optimization decision log; the structured load forecast curve is returned to the front-end interface for visualization through an API interface.

[0032] Preferably, an optimization decision log is generated based on the optimal electricity consumption planning results. This optimization decision log records key decision information during the load optimization and adjustment process, including at least the status indicator of whether load transfer occurred within each forecast period, the power distribution before and after load transfer, and a comparison of electricity costs calculated based on time-of-use pricing. After generating the optimization decision log, the load forecast curve is structured to establish a one-to-one correspondence with the optimization decision log in both time and load dimensions. The structured load forecast curve and the optimization decision log are then stored together in a temporary database to support subsequent data queries and visualization. Subsequently, the structured load forecast curve is read from the temporary database and returned to the front-end interface via an API interface. Upon receiving the load forecast curve data, the front-end combines it with the optimization decision log content for visualization, thereby simultaneously displaying load change trends, load transfer status, and electricity cost changes to improve the understandability and decision support effect of the electricity consumption planning results.

[0033] Furthermore, a prediction period is defined by a preset time period; based on the iteration of the prediction period, a polling prediction based on electricity load is performed to optimize electricity planning and management.

[0034] The system manages the electricity load forecasting process periodically using a preset time period as the forecast cycle. This preset time period is set to a minute-level, hour-level, or daily forecast cycle based on electricity planning requirements, and serves as the time benchmark for load forecasting and planning updates. During system operation, a polling forecasting process based on electricity load is periodically triggered according to the iteration of the forecast cycle. Specifically, at the end of each forecast cycle, the latest electricity load data and external environmental data are automatically acquired, and combined with the forecast results from the previous forecast cycle and the optimized electricity planning results, the load forecaster is invoked again to perform load forecasting and planning calculations. Through this polling forecasting method, the load forecast curve and electricity planning results can be continuously updated over time, thereby achieving dynamic tracking of electricity load changes and continuous optimization management of electricity planning schemes, improving the adaptability and stability of electricity planning during long-term operation.

[0035] In summary, the embodiments of this application have at least the following technical effects: First, a power planning request is sent to the backend API via HTTP. The API receives the planning request data from the frontend and loads the load forecaster. Then, based on the load forecaster, dual-mode planning is performed on the planning request data to determine the load forecast curve. The load forecaster is built through forward and backward training based on a primitive layer, a semantic layer, and an intent layer. The dual modes include coarse planning and precise planning. Finally, the load forecast curve is visualized on the frontend interface via the API. This solves the technical problem in existing power load forecasting models where insufficient coupling between multi-source data and user planning intent leads to poor forecast results that fail to meet actual power planning needs and have poor usability. By constructing a load forecaster and performing dual-mode load forecasting on planning requests, the usability of power load forecasting results is improved.

[0036] Example 2, based on the same inventive concept as the electricity load forecasting method based on multi-source data fusion in the previous examples, such as... Figure 2 As shown, this application provides an electricity load forecasting system based on multi-source data fusion, wherein the system includes: Data sending module 11: Sends a power planning request to the backend API via HTTP protocol. The API interface receives the planning request data sent by the frontend and loads the load forecaster. Load forecasting module 12: Performs dual-mode planning on the planning request data based on the load forecaster to determine the load forecast curve. The load forecaster is built based on the primitive layer, semantic layer, and intent layer through forward and reverse training. The dual modes include coarse planning and precise planning. Data visualization module 13: Visualizes the load forecast curve on the frontend interface through the API interface.

[0037] Furthermore, the data sending module 11 is used to perform the following method: Historical load curves are retrieved, and load elements, the smallest behavioral units, are deconstructed and determined. A causal inference chain based on external strong driving factors and internal load element activation is reconstructed for each load element to determine a first causal view. Association rules are then reconstructed for each load element to determine a second association view, where the association rules are the symbiotic, mutually exclusive, and sequential associations between load elements in terms of time sequence and spatial location. A generative power spectrum is defined using the first causal view and the second association view.

[0038] Furthermore, the load forecasting module 12 is used to perform the following method: A load predictor is built by employing a multi-layer coding approach based on a primitive layer, a semantic layer, and an intent layer, and by performing bidirectional training. The primitive layer is built based on the generative power graph, the semantic layer is built based on load action phrases under primitive sequence combinations, and the intent layer is built by power load intent reasoning based on load action phrases.

[0039] Furthermore, the load forecasting module 12 is used to perform the following method: Forward training based on the primitive layer, semantic layer, and intent layer is performed to generate a first prediction path, wherein the forward training is based on the primitive assembly splicing generation and semantic action intent reasoning based on the generative power spectrum; inverse training based on the primitive layer, semantic layer, and intent layer is performed to generate a second prediction path, wherein the inverse training is based on the load intent reverse generation under semantic trajectory probability, load action phrase decoding, and power time series instantiation.

[0040] Furthermore, the load forecasting module 12 is used to perform the following method: The system acquires planning request data, which is a load curve generation for a preset time period. The load curve generation is modeled as a sequence generation task based on graph constraints. Multi-source data is collected in real time to generate an external environment vector. Based on the planning request data and the external environment vector, short-term prediction is performed in the loaded load forecaster to generate a load forecast curve.

[0041] Furthermore, the load forecasting module 12 is used to perform the following method: By fusing external environment vectors and planning request data, and based on the first prediction path, the first candidate load map is determined by splicing primitive sequences in the generative power map, and the first load semantic trajectory is recursively deduced; based on the second prediction path, the second load semantic trajectory is determined by decoding, instantiated into the second power time series, and the coarse planning result is determined.

[0042] Furthermore, the load forecasting module 12 is used to perform the following method: Based on the rough planning results, the load intention periods and transferable loads with high electricity price elasticity are determined according to the load semantic trajectory. For the load intention periods and transferable loads, cross-time period optimization and redistribution are performed to determine the optimal electricity consumption plan and generate the load forecast curve. The cross-time period is divided into peak, flat, valley and deep valley as the basis for redistribution.

[0043] Furthermore, the data visualization module 13 is used to perform the following methods: An optimization decision log is generated, in which the load transfer status and electricity cost comparison during different time periods are used as generation elements; the load forecast curve is structured and stored in a temporary database along with the optimization decision log; the structured load forecast curve is returned to the front-end interface for visualization through an API interface.

[0044] Furthermore, the data visualization module 13 is used to perform the following methods: The prediction period is defined by a preset time period; based on the iteration of the prediction period, a polling prediction based on electricity load is performed to optimize electricity planning and management.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for predicting electricity load based on multi-source data fusion, characterized in that, The method includes: The power planning request is sent to the backend API via the HTTP protocol. The API interface receives the planning request data sent by the frontend and loads the load forecaster. According to the load forecaster, dual-mode planning is performed on the planning request data to determine the load forecast curve. The load forecaster is constructed based on the primitive layer, semantic layer and intent layer through forward and reverse training. The dual-mode includes coarse planning and precise planning. The load forecast curve can be visualized through the API interface on the front-end interface.

2. The electricity load forecasting method based on multi-source data fusion as described in claim 1, characterized in that, Before loading the load forecaster, the construction process of the load forecaster includes: Retrieve historical load curves and deconstruct them to determine the load element, which is the smallest behavioral unit; The load primitives are reconstructed based on the causal inference chain from external strong driving factors to the activation of internal load primitives to determine the first causal view; The load primitives are reconstructed using association rules to determine a second association view, wherein the association rules are the symbiotic, mutually exclusive, and sequential associations between load primitives in terms of time sequence and spatial location; A generative power map is defined using the first causal view and the second relational view.

3. The electricity load forecasting method based on multi-source data fusion as described in claim 2, characterized in that, A load predictor is built by adopting a multi-layer coding approach based on primitive layer, semantic layer and intent layer and performing bidirectional training. The primitive layer is built based on the generative power graph, the semantic layer is built based on load action phrases under primitive sequence combinations, and the intent layer is built based on power load intent reasoning based on load action phrases.

4. The electricity load forecasting method based on multi-source data fusion as described in claim 3, characterized in that, By performing bidirectional training, including: Perform forward training based on the primitive layer, semantic layer and intent layer to generate a first prediction path, wherein the forward training is based on the primitive assembly splicing generation and semantic action intent reasoning based on the generative power spectrum. Perform inverse training based on the primitive layer, semantic layer and intent layer to generate a second prediction path. The inverse training is based on the reverse generation of load intent under semantic trajectory probability, the decoding of load action phrases and the instantiation of power time series.

5. The electricity load forecasting method based on multi-source data fusion as described in claim 1, characterized in that, Perform dual-mode planning on the planning request data to determine the load forecast curve, including: Obtain planning request data, wherein the planning request data is a load curve generation for a preset time period, and the load curve generation is modeled as a sequence generation task based on graph constraints; Real-time acquisition of multi-source data to generate external environment vectors; Based on the planning request data and the external environment vector, short-term forecasting is performed in the loaded load forecaster to generate a load forecast curve.

6. The electricity load forecasting method based on multi-source data fusion as described in claim 5, characterized in that, Short-term forecasting is performed in the loaded load forecaster, including: By fusing external environment vectors and planning request data, and based on the first prediction path, the first candidate load map is determined by splicing primitive sequences in the generative power map, and the semantic trajectory of the first load is recursively deduced. Based on the second prediction path, the second load semantic trajectory is determined by decoding, instantiated into the second power time series, and the coarse planning result is determined.

7. The electricity load forecasting method based on multi-source data fusion as described in claim 6, characterized in that, After determining the rough planning results, the following are included: Based on the rough planning results, and according to the load semantic trajectory, the intended periods of high electricity price elasticity and the transferable loads are determined; For the intended load period and transferable load, cross-period optimization redistribution is performed to determine the optimal power consumption plan and generate the load forecast curve. The cross-period redistribution is based on the division of peak, flat, valley and deep valley.

8. The electricity load forecasting method based on multi-source data fusion as described in claim 1, characterized in that, The load forecast curve is visualized via an API interface on the front-end interface, including: Generate an optimization decision log, in which the load transfer status and electricity cost comparison for different time periods are used as generation elements; The load forecast curve is structured and stored in a temporary database along with the optimization decision log; The structured load forecast curve is returned to the front-end interface for visualization via the API interface.

9. The electricity load forecasting method based on multi-source data fusion as described in claim 1, characterized in that, The prediction period is defined by a preset time period; Based on the iteration of the forecast cycle, a polling forecast based on electricity load is performed to optimize electricity planning and management.

10. A power load forecasting system based on multi-source data fusion, characterized in that, The system is used to implement the electricity load forecasting method based on multi-source data fusion as described in any one of claims 1-9, the system comprising: Data sending module: Sends electricity planning requests to the backend API via HTTP protocol. The API interface receives planning request data sent by the frontend and loads the load forecaster. Load forecasting module: Based on the load forecaster, performs dual-mode planning on the planning request data to determine the load forecast curve. The load forecaster is constructed based on the primitive layer, semantic layer and intent layer through forward and reverse training. The dual-mode includes coarse planning and precise planning. Data visualization module: The load forecast curve is visualized through the API interface on the front-end interface.

Images