Load precise prediction method and system for virtual power plant multi-source heterogeneous data fusion

Abstract

The application relates to the technical field of accurate load prediction of a virtual power plant, and provides a load accurate prediction method and system for multi-source heterogeneous data fusion of a virtual power plant, so as to solve the problem that the existing technology is difficult to adapt to the accurate load prediction demand of a virtual power plant for a short period. The application collects photovoltaic output, wind power fluctuation data of a virtual power plant, and real-time current and node voltage monitoring data of a distribution network, associates and labels the former two to obtain an associated data set, phase-calibrates the latter two to obtain a calibrated data set, and integrates the two data sets into a multi-source heterogeneous data set; then, fluctuation characteristics of various data are extracted, an XGBoost algorithm is used to screen out a core information feature set affecting load prediction accuracy, and finally, a GRU time sequence prediction model is used to generate a load prediction result in a preset time period, so that the multi-source heterogeneous data of a virtual power plant can be fused, feature screening and time sequence correlation processing are performed, and accurate load prediction in a preset time period can be realized.

Description

Technical Field

[0001] This application relates to the technical field of accurate load forecasting for virtual power plants, and in particular to a method and system for accurate load forecasting by fusing multi-source heterogeneous data from virtual power plants. Background Technology

[0002] Virtual power plants integrate distributed photovoltaic (PV), wind power, and other renewable energy generation units with distribution network systems to achieve coordinated energy dispatch and optimized operation. Their load forecasting requires comprehensive data from renewable energy sources such as PV output and wind power fluctuations, as well as distribution network operation data such as line current and node voltage. These data types are diverse and heterogeneous; accurately integrating such data to improve load forecasting accuracy is a core requirement supporting the efficient absorption of renewable energy and ensuring grid stability in virtual power plants.

[0003] Currently, existing technologies for load forecasting mostly rely on historical load data, combined with single-type auxiliary data such as regional meteorological data, and employ traditional time series models, such as autoregressive moving average models, for forecasting. In terms of data processing, single-source data is typically standardized and directly input into the model. The forecast results are obtained by learning the statistical relationship between historical data and load, focusing on fitting conventional electricity consumption patterns.

[0004] However, the most significant drawback of existing technologies is that they cannot effectively integrate multi-source heterogeneous data from renewable energy output and distribution network operation. They cannot specifically extract the time-period fluctuation characteristics of different data, nor can they screen out key information that has a significant impact on load forecasting. As a result, the impact of dynamic factors such as renewable energy output fluctuations and changes in distribution network status on load is not fully captured, which in turn affects forecast accuracy and makes it difficult to meet the needs of virtual power plants for accurate short-term load forecasting. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for accurate load forecasting by fusing multi-source heterogeneous data from a virtual power plant, in order to solve the problem that existing technologies are difficult to adapt to the accurate short-term load forecasting requirements of virtual power plants.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a method for accurate load forecasting based on the fusion of multi-source heterogeneous data from a virtual power plant, comprising:

[0007] Collect photovoltaic power output data and wind power fluctuation data of new energy sources in the virtual power plant, as well as real-time line current data and node voltage monitoring data of the distribution network;

[0008] The photovoltaic output data and the wind power fluctuation data are correlated and labeled to obtain a correlated dataset. The real-time current data of the line and the node voltage monitoring data are phase calibrated to obtain a calibrated dataset. The correlated dataset and the calibrated dataset are integrated to obtain a multi-source heterogeneous dataset.

[0009] Extract time-period fluctuation information of various types of data from the multi-source heterogeneous dataset to obtain the fluctuation characteristics of various types of data;

[0010] The XGBoost algorithm is used to evaluate the feature importance of the fluctuation characteristics of the various types of data, and obtain the quantitative value of the influence of the fluctuation characteristics of the various types of data on the accuracy of load forecasting results. The fluctuation characteristics with the quantitative value of the influence greater than a preset threshold are selected from the fluctuation characteristics of the various types of data to obtain the core information feature set.

[0011] The GRU time-series forecasting model is used to perform time-series correlation processing on the core information feature set to obtain the time-series correlation rules between different fluctuation features in the core information feature set. Based on the time-series correlation rules, the load forecast results of the virtual power plant within a preset time period are generated.

[0012] Optionally, the XGBoost algorithm is used to evaluate the feature importance of the fluctuation characteristics of the various types of data, obtain a quantified value of the influence of the fluctuation characteristics of the various types of data on the accuracy of load forecasting results, and select fluctuation characteristics whose quantified influence value is greater than a preset threshold from the fluctuation characteristics of the various types of data to obtain a core information feature set, including:

[0013] Based on historical load data and the fluctuation characteristics of the aforementioned data types, the XGBoost algorithm is used to establish a correlation model between fluctuation characteristics and load changes. Based on the correlation model, the contribution of each fluctuation characteristic to load forecasting is analyzed, and the quantitative value of the influence of each fluctuation characteristic is obtained.

[0014] Based on the quantified value of the degree of influence, the fluctuation characteristics of various types of data are sorted to generate a feature importance ranking list;

[0015] Based on a preset feature selection threshold, fluctuating features with feature importance scores higher than the feature selection threshold are selected from the feature importance ranking list to form a core information feature set.

[0016] Optionally, based on historical load data and the fluctuation characteristics of the various types of data, the XGBoost algorithm is used to establish a correlation model between fluctuation characteristics and load changes. Based on the correlation model, the contribution of each fluctuation characteristic to load forecasting is analyzed to obtain a quantitative value of the influence of each fluctuation characteristic, including:

[0017] The XGBoost algorithm is used to combine historical load data with the fluctuation characteristics of the aforementioned data according to their time correspondence to form a training sample set;

[0018] Based on the training sample set, multiple decision trees are constructed to form a correlation model between fluctuation characteristics and load changes;

[0019] Statistical analysis of the correlation model yields the number of times each fluctuation feature is used as a split node and the quantitative value of the improvement in load forecast accuracy brought about by its use as a split node.

[0020] Based on the number of uses and the quantified value of the improvement in load forecast accuracy, the weight coefficient of each fluctuation feature is calculated to obtain the quantified value of the degree of influence.

[0021] Optionally, the step of using the GRU time-series prediction model to perform time-series correlation processing on the core information feature set to obtain the time-series correlation patterns between different fluctuation features in the core information feature set includes:

[0022] The GRU time-series prediction model is used to perform time-series correlation processing on the core information feature set to obtain a time feature sequence;

[0023] Based on the aforementioned time feature sequence, the degree of retention of historical load feature information is controlled by the update gate of the GRU time series prediction model. The update gate retains the corresponding historical load feature information based on the sharp increase in photovoltaic output in the current time series feature sequence. The degree of forgetting of historical load feature information is controlled by the reset gate of the GRU time series prediction model. The reset gate decides to forget the corresponding historical load feature information based on the characteristic that photovoltaic output drops to zero in the current time series feature sequence.

[0024] Based on the outputs of the update gate and the reset gate, and combined with the core information features of the current moment in the core information feature set, the hidden state of the GRU time series prediction model is updated, and the time dependency relationship between photovoltaic power output change features, wind power fluctuation features and distribution network voltage deviation features is captured to form a time series correlation law.

[0025] Optionally, based on the outputs of the update gate and the reset gate, and combined with the core information features of the current moment in the core information feature set, the hidden state of the GRU time-series prediction model is updated, and the time dependency relationship between photovoltaic power output change features, wind power fluctuation features, and distribution network voltage deviation features is captured to form a time-series correlation law, including:

[0026] The output of the update gate is combined with the hidden state of the previous time step to generate candidate retention information;

[0027] The output of the reset gate is fused with the candidate retention information to obtain the candidate state at the current moment;

[0028] The core information features of the current moment are integrated with the candidate states to form a new hidden state;

[0029] During the hidden state update process, the leading lag relationship between the characteristics of photovoltaic power output change and the characteristics of wind power fluctuation is analyzed, and the response mode of distribution network voltage deviation characteristics to photovoltaic power output change and wind power fluctuation is identified.

[0030] The leading lag relationship is integrated with the response pattern to form a time-series correlation law.

[0031] Optionally, the process of performing correlation annotation on the photovoltaic power output data and the wind power fluctuation data to obtain a correlation dataset, and performing phase calibration on the real-time line current data and the node voltage monitoring data to obtain a calibrated dataset, includes:

[0032] Based on the output values ​​at each time point in the photovoltaic output data, the irradiance values ​​for the corresponding time period are matched to form the first binding data of photovoltaic output and irradiance. Based on the power values ​​at each time point in the wind power fluctuation data, the wind speed level range is divided to form the second binding data of wind power and wind speed level.

[0033] The first bound data and the second bound data are combined to obtain the associated dataset;

[0034] Identify the time stamps of each monitoring point in the real-time current data of the line and the node voltage monitoring data. Based on the time stamps, perform phase calibration on the current values ​​in the real-time current data of the line and the voltage values ​​in the node voltage monitoring data to obtain a calibrated dataset.

[0035] Optionally, the step of extracting time-period fluctuation information of various types of data from the multi-source heterogeneous dataset to obtain the fluctuation characteristics of various types of data includes:

[0036] Based on the photovoltaic power output data in the multi-source heterogeneous dataset, the change in power output value between adjacent time points is calculated to obtain the photovoltaic power output change sequence.

[0037] Based on the wind power fluctuation data in the multi-source heterogeneous dataset, the difference between the maximum and minimum power values ​​within a preset time window is calculated to obtain the wind power fluctuation sequence.

[0038] Based on the node voltage monitoring data in the multi-source heterogeneous dataset, the degree of deviation between the voltage value of each monitoring point and the reference voltage is calculated to obtain the voltage deviation sequence;

[0039] The photovoltaic power output change sequence, the wind power fluctuation sequence, and the voltage deviation sequence are integrated to form the fluctuation characteristics of the various types of data.

[0040] Secondly, this application provides a load forecasting system for accurate prediction based on the fusion of multi-source heterogeneous data from a virtual power plant, comprising:

[0041] The data acquisition module is used to collect photovoltaic power output data and wind power fluctuation data of new energy sources in the virtual power plant, as well as real-time line current data and node voltage monitoring data of the distribution network.

[0042] An integration module is used to perform correlation and annotation processing on the photovoltaic output data and the wind power fluctuation data to obtain a correlation dataset, perform phase calibration on the real-time current data of the line and the node voltage monitoring data to obtain a calibrated dataset, and integrate the correlation dataset and the calibrated dataset to obtain a multi-source heterogeneous dataset.

[0043] The extraction module is used to extract time-period fluctuation information of various types of data from the multi-source heterogeneous dataset to obtain the fluctuation characteristics of various types of data;

[0044] The evaluation module is used to evaluate the feature importance of the fluctuation characteristics of the various types of data using the XGBoost algorithm, obtain the quantification value of the influence of the fluctuation characteristics of the various types of data on the accuracy of load forecasting results, and select the fluctuation characteristics with the quantification value of the influence greater than a preset threshold from the fluctuation characteristics of the various types of data to obtain the core information feature set.

[0045] The correlation module is used to perform time-series correlation processing on the core information feature set using the GRU time-series prediction model, to obtain the time-series correlation rules between different fluctuation features in the core information feature set, and to generate the load prediction results of the virtual power plant within a preset time period based on the time-series correlation rules.

[0046] Thirdly, this application provides an electronic device, comprising:

[0047] Memory, used to store computer programs;

[0048] A processor is configured to execute the computer program to implement the steps of the load accurate forecasting method for virtual power plant multi-source heterogeneous data fusion as described in the first aspect above.

[0049] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the load prediction method for multi-source heterogeneous data fusion in a virtual power plant as described in the first aspect above.

[0050] The load forecasting method based on multi-source heterogeneous data fusion for virtual power plants provided in this application collects photovoltaic output data and wind power fluctuation data from new energy sources in the virtual power plant, as well as real-time line current data and node voltage monitoring data from the distribution network. The photovoltaic output data and wind power fluctuation data are correlated and labeled to obtain a correlated dataset. The real-time line current data and node voltage monitoring data are phase-calibrated to obtain a calibrated dataset. The correlated dataset and the calibrated dataset are integrated to obtain a multi-source heterogeneous dataset. Time-period fluctuation information of various data types is extracted from the multi-source heterogeneous dataset to obtain fluctuation characteristics of each data type. The XGBoost algorithm is used to evaluate the feature importance of the fluctuation characteristics of each data type, obtaining a quantified value of the influence of the fluctuation characteristics of each data type on the accuracy of the load forecasting results. Fluctuation characteristics with a quantified influence value greater than a preset threshold are selected from the fluctuation characteristics of each data type to obtain a core information feature set. The GRU time-series forecasting model is used to perform time-series correlation processing on the core information feature set to obtain the time-series correlation rules between different fluctuation characteristics in the core information feature set. Based on the time-series correlation rules, the load forecasting results of the virtual power plant within a preset time period are generated.

[0051] The technical solution of this application has the following beneficial effects:

[0052] This application provides multi-source basic data covering both renewable energy output and distribution network operation for virtual power plant load forecasting by collecting photovoltaic power output data, wind power fluctuation data, real-time line current data, and node voltage monitoring data, ensuring the comprehensiveness and relevance of data sources. By correlating and labeling renewable energy data and performing phase calibration on distribution network data, and integrating them into a multi-source heterogeneous dataset, preliminary standardization of different data types is achieved, resolving the issue of inconsistent data formats and time bases, and providing a unified data foundation for subsequent fusion analysis. Furthermore, by extracting time-period fluctuation information from various data types in the multi-source heterogeneous dataset, fluctuation characteristics are obtained, capturing renewable energy output. By analyzing the dynamic changes in distribution network operation, key dynamic attributes related to load in the data are highlighted, providing targeted input for feature selection. The XGBoost algorithm is used to evaluate the importance of fluctuation characteristics and select the core information feature set, eliminating redundant features with little impact on load forecasting, reducing data dimensionality, ensuring that the features input to the prediction model have high correlation, and improving model processing efficiency. The GRU time series prediction model is used to process the core information feature set and generate prediction results, effectively capturing the time series correlation patterns between different fluctuation characteristics. Combined with dynamic correlation, accurate short-term load forecast results are generated, adapting to the dynamic response requirements of virtual power plants to load changes.

[0053] Furthermore, based on historical load data and the fluctuation characteristics of various types of data, this application uses the XGBoost algorithm to establish a correlation model between fluctuation characteristics and load changes, analyzes the contribution of each fluctuation characteristic in load forecasting to obtain a quantitative value of influence, sorts the fluctuation characteristics based on the quantitative value to generate a ranking list of feature importance, and then selects high-importance fluctuation characteristics from the list according to a preset threshold to form a core information feature set.

[0054] This application establishes a correlation model to accurately quantify the impact of various fluctuation characteristics on load forecasting, and scientifically extracts core features by combining sorting and threshold screening. This ensures high correlation of the core information feature set while reducing interference from redundant information, providing high-quality input for subsequent time series forecasting and improving the accuracy and efficiency of load forecasting. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 A flowchart illustrating a method for accurate load forecasting based on multi-source heterogeneous data fusion from a virtual power plant, provided in an embodiment of this application;

[0057] Figure 2 A schematic diagram illustrating a specific implementation of a method for accurate load forecasting by fusing multi-source heterogeneous data from a virtual power plant, provided in an embodiment of this application;

[0058] Figure 3 This is a schematic diagram of the structure of a virtual power plant multi-source heterogeneous data fusion load prediction system provided in an embodiment of this application;

[0059] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0060] In existing virtual power plant load forecasting, the diverse data types and formats of renewable energy output and distribution network operation make it difficult for current technologies to effectively integrate these scattered and heterogeneous data. They often rely on single data points or simple processing for direct forecasting, failing to capture the fluctuation patterns of different data over time or distinguish which data is more important to the forecast results. This leads to key factors such as renewable energy fluctuations and changes in distribution network status not being fully considered, resulting in forecast accuracy that cannot meet the requirements for efficient virtual power plant dispatching.

[0061] To address this, this application proposes a method for accurate load forecasting based on the fusion of multi-source heterogeneous data from a virtual power plant. This method first collects data from renewable energy sources such as photovoltaic and wind power, as well as distribution network operation data, integrating them into a unified dataset through association labeling and phase calibration. Next, it extracts the time-varying characteristics of various data types and uses the XGBoost algorithm to filter out core features that significantly impact forecasting. Finally, it uses a GRU time-series model to capture the temporal correlation patterns of these core features, generating accurate forecast results. This process not only solves the problem of fusion of multi-source heterogeneous data but also effectively compensates for the shortcomings of existing technologies and improves the accuracy of virtual power plant load forecasting by extracting fluctuation features and filtering key information, thus fully incorporating the impact of renewable energy fluctuations and distribution network changes.

[0062] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0063] The core of this application is to provide a method for accurate load forecasting by fusing multi-source heterogeneous data from a virtual power plant. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0064] S101. Collect photovoltaic power output data and wind power fluctuation data of new energy sources in the virtual power plant, as well as real-time line current data and node voltage monitoring data of the distribution network.

[0065] In the above scheme, photovoltaic power output data refers to the power output data of photovoltaic power plants in different time periods, including the power generation value at different times, which is used to reflect the real-time and changing situation of photovoltaic power generation; wind power fluctuation data refers to the dynamic change data of wind farm power generation over time, including the increase or decrease in power and the frequency of change, which is used to reflect the volatility of wind power output; real-time line current data refers to the real-time monitoring value of the current passing through the distribution network lines, which reflects the load status of the lines; node voltage monitoring data refers to the real-time measurement value of the voltage at each node in the distribution network, which is used to reflect the voltage stability of the distribution network.

[0066] In this application example, firstly, the new energy access points and distribution network monitoring range of the virtual power plant are determined. For photovoltaic power output data, data acquisition terminals are deployed at the inverters of each photovoltaic power station, with a fixed acquisition interval of 15 minutes. The terminals read and record the power output value of the inverters in real time. Next, for wind power fluctuation data, acquisition modules are installed at the wind turbine control cabinets of the wind farm, with a acquisition interval of 10 minutes. The acquisition modules continuously acquire the real-time output power of the wind turbines and record its changes over time. Simultaneously, for real-time line current data, current sensors are installed on the feeder terminal units of each line in the distribution network. The sensors monitor the current value in the line in real time and transmit the data to the acquisition system. Finally, for node voltage monitoring data, voltage sensors are installed at key nodes in the distribution network. The sensors measure the voltage value of the nodes in real time and upload it to the data storage system. By summarizing and integrating the four types of data in the above manner, the data acquisition is completed.

[0067] In practical applications, a virtual power plant connects to two photovoltaic power stations (A and B), one wind farm (C), and a distribution network covering three lines (D, E, and F). To collect photovoltaic output data, a data acquisition terminal is deployed at the inverters of both photovoltaic power stations A and B, set to collect data every 15 minutes. The photovoltaic output data recorded by the terminal is in the format of timestamp + power value. For example, at 8:00, the output of photovoltaic power station A is recorded as 190kW and that of photovoltaic power station B as 210kW; at 8:15, the output of photovoltaic power station A is recorded as 200kW and that of photovoltaic power station B as 220kW. To collect wind power fluctuation data, a data acquisition module was installed at the control cabinets of each of the 10 wind turbines in Wind Farm C. The module was set to collect data every 10 minutes. The wind power data recorded by the module included a timestamp, power value, and power difference between adjacent time periods. For example, at 8:00, the power of turbine 1 was recorded as 260kW and turbine 2 as 250kW; at 8:10, the power of turbine 1 was recorded as 240kW, a decrease of 20kW compared to 8:00; and the power of turbine 2 was recorded as 260kW, an increase of 10kW compared to 8:00. To collect real-time line current data, current sensors were installed at the feeder terminal units of lines D, E, and F. The sensors collected the current value once per second and recorded the average value over one minute. For example, at 8:00, the current of line D was recorded as 150A, line E as 140A, and line F as 160A; at 8:01, the current of line D as 152A, line E as 141A, and line F as 159A. To collect node voltage monitoring data, voltage sensors are installed at three key nodes (G, H, and I) of the distribution network. Each sensor collects a voltage value every second and records the average value over one minute. For example, at 8:00, the voltage at node G is recorded as 10.2 kV, at node H as 10.3 kV, and at node I as 10.4 kV; at 8:01, the voltage at node G is recorded as 10.3 kV, at node H as 10.2 kV, and at node I as 10.5 kV. All the collected data is aggregated via a communication network to the central database of the virtual power plant. This data will be used in subsequent steps for correlation labeling, phase calibration, and the construction of multi-source heterogeneous datasets.

[0068] The aforementioned S101 overall solution collects targeted data on the new energy output of the virtual power plant and the operation data of the distribution network, covering the power generation status of photovoltaic and wind power as well as the current and voltage information of the distribution network. This provides comprehensive and real-time raw data support for subsequent data processing, feature extraction and other steps, ensuring that subsequent analysis can be carried out based on complete basic data and laying a data foundation for improving the accuracy of load forecasting.

[0069] S102. Perform correlation annotation processing on the photovoltaic power output data and the wind power fluctuation data to obtain a correlation dataset. Perform phase calibration on the real-time current data of the line and the node voltage monitoring data to obtain a calibrated dataset. Integrate the correlation dataset and the calibrated dataset to obtain a multi-source heterogeneous dataset.

[0070] Optionally, step S102 involves performing correlation annotation processing on the photovoltaic power output data and the wind power fluctuation data to obtain a correlation dataset, and performing phase calibration on the real-time line current data and the node voltage monitoring data to obtain a calibrated dataset, including:

[0071] Step 1021: Based on the output values ​​at each time point in the photovoltaic output data, match the irradiance values ​​for the corresponding time period to form the first binding data of photovoltaic output and irradiance. Based on the power values ​​at each time point in the wind power fluctuation data, divide the wind speed level range to form the second binding data of wind power and wind speed level.

[0072] Step 1022: Combine the first binding data with the second binding data to obtain the associated dataset.

[0073] Step 1023: Identify the time stamps of each monitoring point in the real-time current data of the line and the node voltage monitoring data. Based on the time stamps, perform phase calibration on the current value in the real-time current data of the line and the voltage value in the node voltage monitoring data to obtain the calibrated dataset.

[0074] In the above scheme, the association annotation process refers to the process of adding relevant environmental parameter information to photovoltaic power output data and wind power fluctuation data to reflect the correlation between new energy output and environmental factors; the first binding data refers to the set of corresponding data of output value at each time point in photovoltaic power output data and irradiance value at the corresponding time period, to reflect the correlation between photovoltaic power generation and light intensity; the second binding data refers to the set of corresponding data of power value at each time point in wind power fluctuation data and corresponding wind speed level range, to reflect the correlation between wind power and wind speed; the association dataset refers to the combination result of the first binding data and the second binding data, which contains the correlation information between new energy output and environmental factors; phase calibration refers to the process of adjusting the time base of line real-time current data and node voltage monitoring data to make the time mark of data from different monitoring points consistent; the calibrated dataset refers to the set of line current and node voltage data after phase calibration, which has a unified time base; the multi-source heterogeneous dataset refers to the integration result of the association dataset and the calibrated dataset, which contains multiple types of association data between new energy and distribution network, for subsequent feature extraction and analysis.

[0075] In this application example, firstly, through step 1021, for photovoltaic power output data, the timestamps of each data are extracted, and the irradiance values ​​of the corresponding timestamps are retrieved from the meteorological data system accessed by the virtual power plant. A one-to-one correspondence is established between the photovoltaic power output values ​​and irradiance values ​​of the same timestamp to form the first binding data. For wind power fluctuation data, the corresponding interval standard of wind speed level and power value is preset. Based on the wind power value of each timestamp, the wind speed level interval to which it belongs is determined to form the second binding data.

[0076] Then, in step 1022, the first bound data and the second bound data are aligned and integrated according to the timestamp. The photovoltaic-irradiance data and the wind power-wind speed level data under the same timestamp are grouped together to obtain the associated dataset.

[0077] Next, in step 1023, the time stamps of each monitoring point in the real-time current data and node voltage monitoring data of the line are extracted, a unified time reference is selected, and the current and voltage values ​​of different monitoring points are adjusted to the time reference so that the same time point contains the corresponding current and voltage data, thus obtaining the calibrated dataset.

[0078] Finally, the associated dataset and the calibrated dataset are merged according to the timestamp. The new energy associated data and distribution network calibration data under the same timestamp are integrated together to obtain a multi-source heterogeneous dataset.

[0079] In practical applications, based on the dataset collected by S101, when processing data from photovoltaic power plant A, the timestamp corresponding to the 190kW output at 8:00 is extracted, and the irradiance at that timestamp is obtained from the meteorological system connected to the virtual power plant, which is 680W / ㎡. These two are then linked to form the first set of bound data. When processing data from photovoltaic power plant B at 8:15, the irradiance at the corresponding timestamp is obtained as 720W / ㎡, forming another set of first-bound data. When processing data from wind farm C, the preset wind speed level classification standard is that level 4 corresponds to a power of 200-250kW, and level 5 corresponds to a power of 251-300kW. The power of wind turbine No. 1 at 8:00, which is 260kW, falls within the level 5 wind speed range, forming the second set of bound data. The power of wind turbine No. 2 at 8:10, which is also 260kW, falls within the level 5 wind speed range, forming another set of second-bound data. These first and second-bound data are then integrated according to their timestamps to obtain a related dataset, which includes new energy-related information for time points such as 8:00, 8:10, and 8:15.

[0080] Next, when processing the distribution network data, the current of 150A at 8:00:02 on line D and the voltage of 10.2kV at 8:00:05 on node G are adjusted to 8:00:00 using whole-minute timestamps as a unified reference to form calibrated data; the current of 141A at 8:01:03 on line E and the voltage of 10.2kV at 8:01:01 on node H are adjusted to 8:01:00 to form another set of calibrated data, and the two are integrated to obtain the calibrated dataset.

[0081] Finally, the associated dataset and the calibrated dataset are merged according to the timestamp to obtain a multi-source heterogeneous dataset. This dataset will be used in S103 to extract time-period fluctuation information of photovoltaic, wind power, line current and node voltage.

[0082] The aforementioned S102 overall solution binds new energy output data with environmental factors through correlation annotation processing, clearly demonstrating the correlation between photovoltaic and wind power output and irradiance and wind speed; it unifies the time base of distribution network data through phase calibration, solving the problem of time asynchrony of data from different monitoring points; and the finally integrated multi-source heterogeneous dataset realizes the organic integration of new energy correlation data and distribution network calibration data, eliminating the heterogeneity of data in the correlation and time dimensions, providing a structurally unified and clearly correlated foundation for subsequent extraction of fluctuation features from the data, and ensuring the effectiveness of subsequent processing.

[0083] S103. Extract time-period fluctuation information of various types of data from the multi-source heterogeneous dataset to obtain the fluctuation characteristics of various types of data.

[0084] Optionally, step S103 involves extracting time-period fluctuation information of various types of data from the multi-source heterogeneous dataset to obtain the fluctuation characteristics of each type of data, including:

[0085] Step 1031: Based on the photovoltaic power output data in the multi-source heterogeneous dataset, calculate the change in power output value between adjacent time points to obtain the photovoltaic power output change sequence.

[0086] Step 1032: Based on the wind power fluctuation data in the multi-source heterogeneous dataset, calculate the difference between the maximum and minimum power values ​​within a preset time window to obtain the wind power fluctuation sequence.

[0087] Step 1033: Based on the node voltage monitoring data in the multi-source heterogeneous dataset, calculate the degree of deviation between the voltage value of each monitoring point and the reference voltage to obtain the voltage deviation sequence.

[0088] Step 1034: Integrate the photovoltaic power output change sequence, the wind power fluctuation sequence, and the voltage deviation sequence to form the fluctuation characteristics of the various types of data.

[0089] In the above scheme, time-period fluctuation information refers to the changes in various types of data in the multi-source heterogeneous dataset within different time periods, including dynamic features such as numerical increases and decreases and range changes; the photovoltaic output change sequence refers to the set of differences in output values ​​at adjacent time points calculated based on photovoltaic output data, used to reflect the real-time trend of photovoltaic power generation; the wind power fluctuation sequence refers to the set of differences between the maximum and minimum wind power values ​​within a preset time window, used to reflect the fluctuation range of wind power; the voltage deviation sequence refers to the set of differences between the voltage values ​​at each time point in the node voltage monitoring data and the reference voltage, used to reflect the stable state changes of the distribution network voltage; the fluctuation feature refers to the integrated result of the photovoltaic output change sequence, wind power fluctuation sequence, and voltage deviation sequence, which centrally reflects the dynamic fluctuation patterns of various types of data, providing a basis for subsequent feature selection.

[0090] In this application example, firstly, through step 1031, photovoltaic power output data is extracted from the multi-source heterogeneous dataset, the data is sorted in chronological order, and the difference between the power output values ​​at two adjacent time points is calculated. These differences are then arranged in chronological order to obtain the photovoltaic power output change sequence.

[0091] Next, in step 1032, wind power fluctuation data is extracted from the multi-source heterogeneous dataset. A fixed time window is set, and the sliding window is used to traverse all the data. The difference between the maximum and minimum power values ​​in each window is calculated, and the differences of all windows are arranged in chronological order to obtain the wind power fluctuation sequence.

[0092] Then, in step 1033, node voltage monitoring data is extracted from the multi-source heterogeneous dataset, the distribution network reference voltage is determined, the difference between the voltage value at each time point and the reference voltage is calculated, and these differences are arranged in chronological order to obtain the voltage deviation sequence.

[0093] Finally, through step 1034, the photovoltaic power output change sequence, wind power fluctuation sequence, and voltage deviation sequence are aligned according to the timestamp, so that the three types of data at the same time point form an associated group, and after integration, the fluctuation characteristics of each type of data are formed.

[0094] In practical applications, based on the multi-source heterogeneous dataset obtained from S102, when processing photovoltaic power output data, photovoltaic power station A outputs 190kW at 8:00 and 200kW at 8:15. The output change at adjacent time points is calculated as 200kW minus 190kW equals 10kW; the output is 200kW at 8:15 and 220kW at 8:30, with the change being 220kW minus 200kW equals 20kW. These change values ​​are then arranged in chronological order to obtain the photovoltaic power output change sequence, including 10kW at 8:15 and 20kW at 8:30.

[0095] Next, when processing wind power fluctuation data, a 30-minute time window was set. The power data of wind turbine No. 1 in wind farm C from 8:00 to 8:30 were 260kW, 240kW, 250kW, and 255kW respectively. The maximum value within this window was 260kW, the minimum value was 240kW, and the fluctuation value was 260kW minus 240kW, which equals 20kW. The power data of wind turbine No. 2 during the same period were 250kW, 260kW, 255kW, and 258kW. The maximum value was 260kW, the minimum value was 250kW, and the fluctuation value was 260kW minus 250kW, which equals 10kW. The wind power fluctuation sequence was obtained by arranging the data according to the window time, including 20kW and 10kW corresponding to 8:30.

[0096] Then, when processing node voltage data, the distribution network reference voltage is set to 10kV. The formula for calculating the voltage deviation is: Voltage Deviation = Monitoring Voltage Value - Reference Voltage Value. Here, the monitoring voltage value is the actual measured voltage of the node, and the reference voltage value is the standard distribution network voltage of 10kV. For example, if the monitoring voltage of node G at 8:00 is 10.2kV, the deviation value is 10.2kV minus 10kV, which equals 0.2kV; if the monitoring voltage at 8:01 is 10.3kV, the deviation value is 10.3kV minus 10kV, which equals 0.3kV. The voltage deviation sequence is obtained by arranging the data according to time, including 0.2kV for 8:00 and 0.3kV for 8:01.

[0097] Finally, the three sequences are integrated according to time alignment to form a fluctuation feature, which will be evaluated for feature importance in S104 using the XGBoost algorithm.

[0098] The aforementioned S103 overall scheme extracts real-time changes in photovoltaic output, window fluctuation range of wind power, and benchmark deviation of node voltage. It transforms the static values ​​in the multi-source heterogeneous data into fluctuation features that reflect dynamic changes, thus centrally reflecting the fluctuation trends of various data over time. This provides targeted input for subsequent screening of key features for load forecasting, helps to highlight the impact of new energy output fluctuations and distribution network status changes on load, and lays a feature foundation for improving the accuracy of load forecasting.

[0099] S104. Using the XGBoost algorithm, the feature importance of the fluctuation characteristics of the various types of data is evaluated to obtain the quantification value of the influence of the fluctuation characteristics of the various types of data on the accuracy of the load forecast result. The fluctuation characteristics with the quantification value of the influence greater than the preset threshold are selected from the fluctuation characteristics of the various types of data to obtain the core information feature set.

[0100] Optionally, in step S104, the XGBoost algorithm is used to evaluate the feature importance of the fluctuation characteristics of the various types of data, obtain a quantified value of the influence of the fluctuation characteristics of the various types of data on the accuracy of load forecasting results, and select fluctuation characteristics whose quantified influence value is greater than a preset threshold from the fluctuation characteristics of the various types of data to obtain the core information feature set, such as... Figure 2 As shown, it includes:

[0101] Step 1041: Based on the historical load data and the fluctuation characteristics of the various types of data, the XGBoost algorithm is used to establish a correlation model between fluctuation characteristics and load changes. Based on the correlation model, the contribution of each fluctuation characteristic in load forecasting is analyzed, and the quantitative value of the influence of each fluctuation characteristic is obtained.

[0102] Step 1041 may specifically include the following steps: using the XGBoost algorithm, combining historical load data with the fluctuation characteristics of the various types of data according to the time correspondence to form a training sample set; based on the training sample set, constructing multiple decision trees to form a correlation model between fluctuation characteristics and load changes; statistically analyzing the correlation model to obtain the number of times each fluctuation characteristic is used as a split node and the quantitative value of the improvement in load prediction accuracy brought about by being used as a split node; calculating the weight coefficient of each fluctuation characteristic according to the number of times it is used and the quantitative value of the improvement in load prediction accuracy to obtain the quantitative value of the degree of influence.

[0103] Step 1042: Based on the quantified value of the degree of influence, sort the fluctuation characteristics of various types of data to generate a feature importance ranking list.

[0104] Step 1043: Based on the preset feature selection threshold, select the fluctuating features whose feature importance scores are higher than the feature selection threshold from the feature importance ranking list to form a core information feature set.

[0105] In the above scheme, the XGBoost algorithm refers to an ensemble learning algorithm based on decision trees, used for feature importance assessment and model construction; feature importance assessment refers to the process of analyzing the impact of various fluctuation features on load forecasting results; the influence degree quantification value refers to the numerical value that measures the contribution of each fluctuation feature to the accuracy of load forecasting, and the larger the value, the more critical the feature; the preset threshold refers to the critical value set by humans to screen important features; the core information feature set refers to the set of fluctuation features that have a significant impact on load forecasting after screening; the training sample set refers to the combination data of historical load data and fluctuation features according to time, used for model training; the correlation model refers to the model that reflects the correlation between fluctuation features and load changes; the decision tree refers to the basic building block of the model, which divides the data through split nodes; the split node refers to the feature point in the decision tree used to divide the data; the weight coefficient refers to the parameter that calculates the influence degree by comprehensively considering the number of times it is used and the degree of accuracy improvement; the feature importance ranking list refers to the list of fluctuation features ranked according to the influence degree quantification value.

[0106] In this application example, firstly, through step 1041, historical load data and the fluctuation characteristics of various data types are combined according to their time correspondence to form a training sample set. Based on this training sample set, multiple decision trees are constructed using the XGBoost algorithm. Each decision tree selects a fluctuation characteristic as a split node to divide the data. The multiple decision trees together constitute a correlation model between fluctuation characteristics and load changes. Subsequently, the correlation model is statistically analyzed to obtain the number of times each fluctuation characteristic is used as a split node, and the quantitative value of the improvement in load prediction accuracy each time it is used as a split node. The weight coefficient of each fluctuation characteristic is calculated by combining the number of times it is used and the quantitative value of the improvement, thereby obtaining the quantitative value of the degree of influence.

[0107] Then, in step 1042, all fluctuation features are sorted from largest to smallest according to their impact quantification values ​​to generate a feature importance ranking list.

[0108] Then, in step 1043, a preset feature selection threshold is set, and fluctuating features with a quantified value of influence greater than the threshold are selected from the feature importance ranking list. These features are then integrated to form a core information feature set.

[0109] In practical applications, based on the fluctuation characteristics obtained from S103, including the photovoltaic power output change sequence, wind power fluctuation sequence, and voltage deviation sequence, historical load data of the virtual power plant over the past 6 months are collected. The historical load data from 8:00 to 8:30, such as 500kW load at 8:00, 520kW load at 8:15, and 550kW load at 8:30, are combined with the corresponding fluctuation characteristics, such as 10kW change in photovoltaic power at 8:15 and 20kW change in photovoltaic power at 8:30, 20kW fluctuation in wind power at 8:30 and 10kW respectively, and voltage deviations of 0.2kV at 8:00 and 0.3kV at 8:01, according to the time correspondence to form a training sample set.

[0110] Next, the XGBoost algorithm was used to construct a correlation model between fluctuation characteristics and load changes. Statistics showed that a 20kW change in photovoltaic power was used as a splitting node 12 times, with an average quantified improvement in prediction accuracy of 0.1 per instance; a 20kW change in wind power was used 8 times, with an average improvement of 0.08; and a 0.3kV voltage deviation was used 5 times, with an average improvement of 0.05. The quantified impact value was then calculated using the formula: Quantified Impact Value = α × Number of Uses + β × Average Quantified Improvement, where α and β are the weighting coefficients for the number of uses and the average quantified improvement, respectively. Here, β is set to 0.6 and β is set to 0.4. The calculated quantified impact value for a 20kW change in photovoltaic power is 0.6 × 12 + 0.4 × 0.1 = 7.24; for a 20kW change in wind power, it is 0.6 × 8 + 0.4 × 0.08 = 4.83; and for a 0.3kV voltage deviation, it is 0.6 × 5 + 0.4 × 0.05 = 3.02.

[0111] Then, the features are sorted from largest to smallest according to their quantified impact values, generating a feature importance ranking list as follows: PV change 20kW, wind power fluctuation 20kW, voltage deviation 0.3kV, etc. A preset threshold of 4 is set, and PV change 20kW and wind power fluctuation 20kW with quantified values ​​greater than 4 are selected to form the core information feature set. This core information feature set will be used in S105 for time series correlation processing of the GRU time series prediction model.

[0112] The S104 overall scheme described above uses the XGBoost algorithm to assess the importance of fluctuation features, quantifies the impact of each feature on load forecasting, and then filters out the core information feature set through a preset threshold. This effectively eliminates redundant features that contribute little to the forecast, reduces the amount of data processed by the subsequent model, and retains key influencing factors, making the features input into the forecasting model more targeted. This provides a high-quality feature foundation for improving the accuracy and efficiency of load forecasting.

[0113] S105. Using the GRU time-series prediction model, perform time-series correlation processing on the core information feature set to obtain the time-series correlation rules between different fluctuation features in the core information feature set. Based on the time-series correlation rules, generate the load prediction results of the virtual power plant within a preset time period.

[0114] Optionally, step S105 involves using a GRU time-series prediction model to perform time-series correlation processing on the core information feature set to obtain the time-series correlation patterns between different fluctuation features in the core information feature set, including:

[0115] Step 1051: Using the GRU time series prediction model, perform time series correlation processing on the core information feature set to obtain a time feature sequence.

[0116] Step 1052: Based on the time feature sequence, the degree of retention of historical load feature information is controlled by the update gate of the GRU time series prediction model. The update gate retains the corresponding historical load feature information according to the photovoltaic power output surge feature in the current time series feature sequence. The degree of forgetting of historical load feature information is controlled by the reset gate in the GRU time series prediction model. The reset gate decides to forget the corresponding historical load feature information according to the photovoltaic power output dropping to zero feature in the current time series feature sequence.

[0117] Step 1053: Based on the output of the update gate and the reset gate, and combined with the core information features of the current moment in the core information feature set, update the hidden state of the GRU time series prediction model, and capture the time dependency relationship between photovoltaic power output change features, wind power fluctuation features and distribution network voltage deviation features to form a time series correlation law.

[0118] Step 1053 may specifically include the following steps: combining the output of the update gate with the hidden state of the previous moment to generate candidate retention information; fusing the output of the reset gate with the candidate retention information to obtain the candidate state of the current moment; integrating the core information features of the current moment with the candidate state to form a new hidden state; during the hidden state update process, analyzing the leading lag relationship between photovoltaic power output change features and wind power fluctuation features, and simultaneously identifying the response mode of distribution network voltage deviation features to photovoltaic power output change and wind power fluctuation; integrating the leading lag relationship with the response mode to form a time-series correlation law, wherein the update stops when all features at all time points in the core information feature set have been processed during the hidden state update process. Specifically, the GRU time-series prediction model processes each feature in the core information feature set sequentially, updating the hidden state once for each feature processed, until the core information features of the last time point in the feature set are traversed. At this point, the hidden state update process stops, so that a time-series correlation law can be formed based on the final hidden state and the relationships captured in the intermediate process.

[0119] In the above scheme, the GRU time-series forecasting model refers to a recurrent neural network model suitable for processing time-series data, capable of capturing the dependencies of data changes over time; the core information feature set refers to the set of fluctuation features that significantly affect load forecasting after screening; time-series correlation processing refers to the process of analyzing the interrelationships of different fluctuation features in the core information feature set over time; time-series correlation law refers to the dependencies and response patterns of different fluctuation features in the time dimension; the preset time period refers to the pre-set time range for load forecasting; the load forecasting result refers to the load value within the preset time period output by the model; the time feature sequence refers to the sequence formed by arranging the core information feature set in chronological order; the update gate refers to the module in the GRU model used to control the proportion of historical load feature information retained; the reset gate refers to the module in the GRU model used to control the proportion of historical load feature information forgotten; the hidden state refers to the variables in the GRU model that store historical information; the leading lag relationship refers to the temporal order of influence between photovoltaic power output change characteristics and wind power fluctuation characteristics; and the response pattern refers to the reaction law of distribution network voltage deviation characteristics to photovoltaic power output change and wind power fluctuation.

[0120] In this application example, firstly, through step 1051, the core information feature set is input into the GRU time series prediction model in chronological order. The model performs time series encoding on the features and arranges the scattered features in chronological order to obtain a time feature sequence.

[0121] Subsequently, based on this time feature sequence, in step 1052, the update gate of the GRU time series prediction model analyzes the photovoltaic output surge feature in the current sequence, calculates the retention weight according to the specific situation of the feature, and controls the degree of retention of historical load feature information; at the same time, the reset gate of the model analyzes the photovoltaic output drop to zero feature in the current sequence, calculates the forgetting weight according to the specific situation of the feature, and controls the degree of forgetting of historical load feature information.

[0122] Next, in step 1053, the output of the update gate is combined with the hidden state of the previous moment to generate candidate retention information; the output of the reset gate is fused with the candidate retention information to obtain the candidate state of the current moment; the core information features of the current moment are integrated with the candidate state to form a new hidden state; during the hidden state update process, the leading lag relationship between the photovoltaic power output change characteristics and the wind power fluctuation characteristics is analyzed, the response mode of the distribution network voltage deviation characteristics to the former two is identified, and these relationships and modes are integrated to form a time-series correlation law.

[0123] Finally, based on this temporal correlation pattern, the load forecast results of the virtual power plant within the preset time period are generated.

[0124] In practical applications, the core information feature set obtained based on S104 includes a 20kW change in photovoltaic power at 8:30, a 20kW fluctuation in wind power at 8:30, a 15kW change in photovoltaic power at 8:45, and an 18kW fluctuation in wind power at 8:45. These features are organized chronologically to form a time feature sequence: 8:30 corresponds to 20kW of photovoltaic power and 20kW of wind power, and 8:45 corresponds to 15kW of photovoltaic power and 18kW of wind power.

[0125] Next, the time feature sequence is input into the GRU time series prediction model. The update gate first detects features in the sequence and finds that the 20kW change in photovoltaic power at 8:30 is a steep increase. Combining this with the influence of similar features on the load in historical data, a retention weight of 0.8 is calculated. This means the model will retain 80% of the historical load feature information, such as the relevant features of the 520kW load at 8:15. Meanwhile, the reset gate does not detect the feature of photovoltaic output dropping to zero, and a forgetting weight of 0.1 is calculated, meaning only 10% of the historical information is forgotten. Then, the 0.8 output of the update gate is multiplied by the hidden state containing the load information of 8:15 from the previous time step to obtain the candidate retention information. Finally, the 0.1 output of the reset gate is fused with the candidate retention information to obtain the candidate state for the current time step.

[0126] Subsequently, the core information features from 8:30 are integrated with the candidate states, and a new hidden state is formed by processing with the tanh activation function. The formula for the tanh activation function is: ,in, The hyperbolic tangent function is a commonly used activation function for nonlinear data transformation. `x` is the input variable of this function, specifically the value obtained after integrating the candidate state with the core information features in this scenario. `e` is the natural constant, approximately 2.71828, a fundamental constant for exponential operations. The 8:45 feature is processed in the same way to update the hidden state. During the hidden state update process, analysis revealed that 3 minutes after the 20kW photovoltaic change, the wind power fluctuation stabilized at 20kW, forming a leading-lag relationship between the two. Simultaneously, 2 minutes after the 20kW photovoltaic change, the distribution network voltage deviation increased from 0.2kV to 0.3kV, forming a voltage deviation response pattern to photovoltaic changes.

[0127] Finally, by integrating these relationships and patterns, a temporal correlation law is obtained. Based on this temporal correlation law, the GRU temporal forecasting model generates load forecast results for the virtual power plant during the preset time period of 9:00-9:30. These results will be used for real-time energy dispatching decisions of the virtual power plant.

[0128] The aforementioned S105 overall solution processes the core information feature set through the GRU time-series prediction model, and uses update gates and reset gates to selectively retain or forget historical information, accurately capturing the time-series correlation patterns between photovoltaic, wind power and distribution network characteristics. This enables the generated load prediction results to fully reflect the dynamic impact of various fluctuation characteristics, improves the accuracy and timeliness of virtual power plant load prediction, and provides a reliable basis for energy dispatch decisions.

[0129] The following is a complete example for steps S101 to S105. A virtual power plant is connected to photovoltaic power station A, photovoltaic power station B, wind farm C, and distribution network system covering three lines D, E, and F. Its load forecasting process is as follows.

[0130] First, a 15-minute data acquisition interval was set at the inverters of photovoltaic power stations A and B. At 8:00, the output of power station A was 190kW and that of power station B was 210kW. At 8:15, the output of power station A was 200kW and that of power station B was 220kW. At the control cabinets of the 10 wind turbines in wind farm C, a 10-minute data acquisition interval was set. At 8:00, the power of wind turbine No. 1 was 260kW and that of wind turbine No. 2 was 250kW. At 8:10, the power of wind turbine No. 1 was 240kW and that of wind turbine No. 2 was 260kW. Current sensors were installed at the feeder terminal units of lines D, E, and F, collecting electricity data once per second. The current is averaged over one minute. At 8:00, the current of line D is recorded as 150A and the current of line E as 140A. At 8:01, the current of line D is recorded as 152A and the current of line E as 141A. Voltage sensors are installed at nodes G, H, and I of the distribution network. The voltage is collected once per second and averaged over one minute. At 8:00, the voltage of node G is recorded as 10.2kV and the voltage of node H as 10.3kV. At 8:01, the voltage of node G is recorded as 10.3kV and the voltage of node H as 10.2kV. All collected data are aggregated to the central database of the virtual power plant through the communication network. These data will be associated, labeled, and calibrated in the future.

[0131] Next, for the photovoltaic power output data, the corresponding irradiance at the time was obtained from the meteorological system connected to the virtual power plant. The irradiance at 8:00 was 680W / ㎡, which was bound to power station A (190kW) and power station B (210kW) to form the first set of bound data. The irradiance at 8:15 was 720W / ㎡, which was bound to power station A (200kW) and power station B (220kW) to form another set of first bound data. For the wind power fluctuation data, the preset wind speed level classification standard is that level 4 corresponds to a power of 200-250kW and level 5 corresponds to a power of 251-300kW. At 8:00, wind turbine No. 1 (260kW) belongs to level 5 and wind turbine No. 2 (250kW) belongs to level 4. At 8:10, wind turbine No. 1 (240kW) belongs to level 4 and wind turbine No. 2 (260kW) belongs to level 5, forming the second set of bound data. The first and second bound data were integrated by time to obtain the associated dataset. For distribution network data, whole-minute timestamps were selected as a unified benchmark. The 150A of line D at 8:00:02 and the 10.2kV of node G at 8:00:05 were calibrated to 8:00:00, and the 141A of line E at 8:01:03 and the 10.2kV of node H at 8:01:01 were calibrated to 8:01:00. The calibrated dataset was then integrated. The associated dataset and the calibrated dataset were merged by time to form a multi-source heterogeneous dataset, which will be used to extract fluctuation characteristics in the future.

[0132] Then, photovoltaic (PV) output data was extracted from the multi-source heterogeneous dataset, and the output changes at adjacent time points were calculated: the difference between 8:15 and 8:00 at power station A was 200kW - 190kW = 10kW, and the difference between 8:15 and 8:00 at power station B was 220kW - 210kW = 10kW. These were arranged according to time to form a PV output change sequence. Wind power fluctuation data was extracted, and a 30-minute time window was set. The power data of wind turbine No. 1 at wind farm C from 8:00 to 8:30 were 260kW, 240kW, 250kW, and 255kW, with a difference of 20kW between the maximum value of 260kW and the minimum value of 240kW. The data of wind turbine No. 2 during the same period were 250kW, 260kW, 255kW, and 258kW, with a difference of 10kW between the maximum value of 260kW and the minimum value of 250kW. These were arranged according to the time window to form a wind power fluctuation sequence. Extract node voltage data, set the distribution network reference voltage to 10kV, and calculate the voltage deviation: the difference between the voltage of node G at 8:00 (10.2kV) and the reference voltage is 10.2kV-10kV=0.2kV, and the difference between the voltage of node G at 8:01 (10.3kV) and the reference voltage is 10.3kV-10kV=0.3kV. Arrange the voltage deviations in time to form a voltage deviation sequence. Align and integrate the three sequences in time to obtain the fluctuation characteristics, which will be used for feature importance assessment later.

[0133] Next, historical load data from the virtual power plant over the past six months was collected, such as 500kW load at 8:00, 520kW load at 8:15, and 550kW load at 8:30. This data was then combined with the obtained fluctuation characteristics according to time to form a training sample set. Subsequently, the XGBoost algorithm was used to construct a correlation model. Statistics showed that a 10kW change in photovoltaic power was used as a splitting node 12 times, with an average quantified improvement in prediction accuracy of 0.1 per instance; a 20kW fluctuation in wind power was used 8 times, with an average improvement of 0.08; and a 0.3kV voltage deviation was used 5 times, with an average improvement of 0.05. The quantified value of the impact was then calculated using the formula: Quantified Impact Value = α × Number of Uses + β × Average Quantified Improvement Value, where α and β are the weighting coefficients for the number of uses and the average quantified improvement value, respectively. Here, α = 0.6 and β = 0.4. The calculated impact values ​​for a 10kW change in photovoltaic power are 0.6×12+0.4×0.1=7.24; for a 20kW fluctuation in wind power, it is 0.6×8+0.4×0.08=4.83; and for a 0.3kV voltage deviation, it is 0.6×5+0.4×0.05=3.02. A feature importance ranking list is then generated by sorting the quantified values ​​from largest to smallest. A preset threshold of 4 is set, and the 10kW change in photovoltaic power and the 20kW fluctuation in wind power are selected to form a core information feature set. This feature set will be used in S105 for processing the GRU time-series prediction model.

[0134] Finally, the core information feature set is organized into a sequence according to time, such as 8:15 corresponding to a 10kW change in photovoltaic power and a 20kW fluctuation in wind power. This sequence is input into the GRU time-series prediction model. The update gate detects the 10kW change in photovoltaic power as a steep increase feature, and calculates a retention weight of 0.8, meaning 80% of the historical load feature information is retained, such as the relevant features of a 500kW load at 8:00. The reset gate does not detect the feature of photovoltaic power output dropping to zero, and calculates a forgetting weight of 0.1, meaning only 10% of the historical information is forgotten. The 0.8 output from the update gate is multiplied by the hidden state containing the 8:00 load information from the previous time step to obtain candidate retention information; the 0.1 output from the reset gate is then fused with the candidate retention information to obtain the candidate state for the current time step. Subsequently, the core information features of 8:15 are integrated with the candidate state, and processed using the tanh activation function to form a new hidden state. The formula for the tanh activation function is... Where x is the value after integrating the candidate state and core information features, and e is the natural constant (approximately 2.71828). During the hidden state update process, it was found that 3 minutes after the photovoltaic change of 10kW, the wind power fluctuation stabilized at 20kW, forming a leading-lag relationship; simultaneously, 2 minutes after the photovoltaic change of 10kW, the distribution network voltage deviation increased from 0.2kV to 0.3kV, forming a response pattern. Integrating these, a time-series correlation pattern was obtained. Based on this pattern, the model generates the load forecast results for the virtual power plant during the preset time period of 9:00-9:30, which will be used for the real-time energy dispatch decision of the virtual power plant.

[0135] Figure 3 This is a schematic diagram illustrating a specific implementation of a virtual power plant multi-source heterogeneous data fusion load forecasting system provided in this application embodiment. (Refer to...) Figure 3 The system may include:

[0136] The data acquisition module 31 is used to collect photovoltaic power output data and wind power fluctuation data of new energy sources in the virtual power plant, as well as real-time line current data and node voltage monitoring data of the distribution network.

[0137] The integration module 32 is used to perform correlation annotation processing on the photovoltaic power output data and the wind power fluctuation data to obtain a correlation dataset, perform phase calibration on the real-time current data of the line and the node voltage monitoring data to obtain a calibrated dataset, and integrate the correlation dataset and the calibrated dataset to obtain a multi-source heterogeneous dataset.

[0138] Extraction module 33 is used to extract time-period fluctuation information of various types of data from the multi-source heterogeneous dataset to obtain the fluctuation characteristics of various types of data.

[0139] Evaluation module 34 is used to evaluate the feature importance of the fluctuation characteristics of the various types of data using the XGBoost algorithm, obtain the quantification value of the influence of the fluctuation characteristics of the various types of data on the accuracy of load forecasting results, select the fluctuation characteristics with the quantification value of the influence greater than a preset threshold from the fluctuation characteristics of the various types of data, and obtain the core information feature set.

[0140] The association module 35 is used to perform time-series association processing on the core information feature set using the GRU time-series prediction model, to obtain the time-series association rules between different fluctuation features in the core information feature set, and to generate the load prediction results of the virtual power plant within a preset time period based on the time-series association rules.

[0141] The virtual power plant multi-source heterogeneous data fusion load accurate prediction system of this application embodiment is used to implement the aforementioned virtual power plant multi-source heterogeneous data fusion load accurate prediction method. Therefore, the specific implementation of the virtual power plant multi-source heterogeneous data fusion load accurate prediction system can be seen in the embodiment section of the virtual power plant multi-source heterogeneous data fusion load accurate prediction method above. The specific implementation can be referred to the description of the corresponding embodiment, and will not be repeated here.

[0142] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of the above-described method for accurate load forecasting by fusing multi-source heterogeneous data from a virtual power plant.

[0143] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for accurate load forecasting by fusing multi-source heterogeneous data from a virtual power plant.

[0144] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0145] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the method for accurate load forecasting by fusing multi-source heterogeneous data from a virtual power plant.

[0146] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0147] The above provides a detailed description of the method and system for accurate load forecasting based on multi-source heterogeneous data fusion in a virtual power plant, as provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for accurate load forecasting based on the fusion of multi-source heterogeneous data from a virtual power plant, characterized in that, include: Collect photovoltaic power output data and wind power fluctuation data of new energy sources in the virtual power plant, as well as real-time line current data and node voltage monitoring data of the distribution network; The photovoltaic output data and the wind power fluctuation data are correlated and labeled to obtain a correlated dataset. The real-time current data of the line and the node voltage monitoring data are phase calibrated to obtain a calibrated dataset. The correlated dataset and the calibrated dataset are integrated to obtain a multi-source heterogeneous dataset. Extract time-period fluctuation information of various types of data from the multi-source heterogeneous dataset to obtain the fluctuation characteristics of various types of data; The XGBoost algorithm is used to evaluate the feature importance of the fluctuation characteristics of the various types of data, and obtain the quantitative value of the influence of the fluctuation characteristics of the various types of data on the accuracy of load forecasting results. The fluctuation characteristics with the quantitative value of the influence greater than a preset threshold are selected from the fluctuation characteristics of the various types of data to obtain the core information feature set. The GRU time-series forecasting model is used to perform time-series correlation processing on the core information feature set to obtain the time-series correlation rules between different fluctuation features in the core information feature set. Based on the time-series correlation rules, the load forecast results of the virtual power plant within a preset time period are generated. The XGBoost algorithm is used to evaluate the feature importance of the fluctuation characteristics of various data types, obtaining a quantified value of the impact of the fluctuation characteristics of various data types on the accuracy of load forecasting results. Fluctuation characteristics with a quantified impact value greater than a preset threshold are selected from the fluctuation characteristics of various data types to obtain a core information feature set, including: Based on historical load data and the fluctuation characteristics of the aforementioned data types, the XGBoost algorithm is used to establish a correlation model between fluctuation characteristics and load changes. Based on the correlation model, the contribution of each fluctuation characteristic to load forecasting is analyzed, and the quantitative value of the influence of each fluctuation characteristic is obtained. Based on the quantified value of the degree of influence, the fluctuation characteristics of various types of data are sorted to generate a feature importance ranking list; Based on a preset feature selection threshold, fluctuating features with feature importance scores higher than the feature selection threshold are selected from the feature importance ranking list to form a core information feature set.

2. The method for accurate load forecasting based on multi-source heterogeneous data fusion in a virtual power plant according to claim 1, characterized in that, Based on historical load data and the fluctuation characteristics of the various data types, the XGBoost algorithm is used to establish a correlation model between fluctuation characteristics and load changes. Based on the correlation model, the contribution of each fluctuation characteristic to load forecasting is analyzed, and a quantitative value of the influence of each fluctuation characteristic is obtained, including: The XGBoost algorithm is used to combine historical load data with the fluctuation characteristics of the aforementioned data according to their time correspondence to form a training sample set; Based on the training sample set, multiple decision trees are constructed to form a correlation model between fluctuation characteristics and load changes; Statistical analysis of the correlation model yields the number of times each fluctuation feature is used as a split node and a quantitative value for the improvement in load forecast accuracy when used as a split node. Based on the number of uses and the quantified value of the improvement in load forecast accuracy, the weight coefficient of each fluctuation feature is calculated to obtain the quantified value of the degree of influence.

3. The method for accurate load forecasting based on multi-source heterogeneous data fusion in a virtual power plant according to claim 1, characterized in that, The GRU time-series prediction model is used to perform time-series correlation processing on the core information feature set to obtain the time-series correlation patterns between different fluctuation features in the core information feature set, including: The GRU time-series prediction model is used to perform time-series correlation processing on the core information feature set to obtain a time feature sequence; Based on the aforementioned time feature sequence, the degree of retention of historical load feature information is controlled by the update gate of the GRU time series prediction model. The update gate retains the corresponding historical load feature information based on the sharp increase in photovoltaic output in the current time series feature sequence. The degree of forgetting of historical load feature information is controlled by the reset gate of the GRU time series prediction model. The reset gate decides to forget the corresponding historical load feature information based on the characteristic that photovoltaic output drops to zero in the current time series feature sequence. Based on the outputs of the update gate and the reset gate, and combined with the core information features of the current moment in the core information feature set, the hidden state of the GRU time series prediction model is updated, and the time dependency relationship between photovoltaic power output change features, wind power fluctuation features and distribution network voltage deviation features is captured to form a time series correlation law.

4. The method for accurate load forecasting based on multi-source heterogeneous data fusion in a virtual power plant according to claim 3, characterized in that, Based on the outputs of the update gate and reset gate, and combined with the core information features of the current moment in the core information feature set, the hidden state of the GRU time-series prediction model is updated, and the time dependency relationship between photovoltaic power output change features, wind power fluctuation features, and distribution network voltage deviation features is captured to form a time-series correlation law, including: The output of the update gate is combined with the hidden state of the previous time step to generate candidate retention information; The output of the reset gate is fused with the candidate retention information to obtain the candidate state at the current moment; The core information features of the current moment are integrated with the candidate states to form a new hidden state; During the hidden state update process, the leading lag relationship between the characteristics of photovoltaic power output change and the characteristics of wind power fluctuation is analyzed, and the response mode of distribution network voltage deviation characteristics to photovoltaic power output change and wind power fluctuation is identified. The leading lag relationship is integrated with the response pattern to form a time-series correlation law.

5. The method for accurate load forecasting based on multi-source heterogeneous data fusion in a virtual power plant according to claim 1, characterized in that, The process involves correlating and labeling the photovoltaic power output data and the wind power fluctuation data to obtain a correlated dataset, and performing phase calibration on the real-time line current data and the node voltage monitoring data to obtain a calibrated dataset, including: Based on the output values ​​at each time point in the photovoltaic output data, the irradiance values ​​for the corresponding time period are matched to form the first binding data of photovoltaic output and irradiance. Based on the power values ​​at each time point in the wind power fluctuation data, the wind speed level range is divided to form the second binding data of wind power and wind speed level. The first bound data and the second bound data are combined to obtain the associated dataset; Identify the time stamps of each monitoring point in the real-time current data of the line and the node voltage monitoring data. Based on the time stamps, perform phase calibration on the current values ​​in the real-time current data of the line and the voltage values ​​in the node voltage monitoring data to obtain a calibrated dataset.

6. The method for accurate load forecasting based on multi-source heterogeneous data fusion in a virtual power plant according to claim 1, characterized in that, The step of extracting time-period fluctuation information of various types of data from the multi-source heterogeneous dataset to obtain the fluctuation characteristics of various types of data includes: Based on the photovoltaic power output data in the multi-source heterogeneous dataset, the change in power output value between adjacent time points is calculated to obtain the photovoltaic power output change sequence. Based on the wind power fluctuation data in the multi-source heterogeneous dataset, the difference between the maximum and minimum power values ​​within a preset time window is calculated to obtain the wind power fluctuation sequence. Based on the node voltage monitoring data in the multi-source heterogeneous dataset, the degree of deviation between the voltage value of each monitoring point and the reference voltage is calculated to obtain the voltage deviation sequence; The photovoltaic power output change sequence, the wind power fluctuation sequence, and the voltage deviation sequence are integrated to form the fluctuation characteristics of the various types of data.

7. A load forecasting system based on the fusion of multi-source heterogeneous data from a virtual power plant, characterized in that, A method for accurate load forecasting by fusing multi-source heterogeneous data from a virtual power plant as described in any one of claims 1 to 6 includes: The data acquisition module is used to collect photovoltaic power output data and wind power fluctuation data of new energy sources in the virtual power plant, as well as real-time line current data and node voltage monitoring data of the distribution network. An integration module is used to perform correlation and annotation processing on the photovoltaic output data and the wind power fluctuation data to obtain a correlation dataset, perform phase calibration on the real-time current data of the line and the node voltage monitoring data to obtain a calibrated dataset, and integrate the correlation dataset and the calibrated dataset to obtain a multi-source heterogeneous dataset. The extraction module is used to extract time-period fluctuation information of various types of data from the multi-source heterogeneous dataset to obtain the fluctuation characteristics of various types of data; The evaluation module is used to evaluate the feature importance of the fluctuation characteristics of the various types of data using the XGBoost algorithm, obtain the quantification value of the influence of the fluctuation characteristics of the various types of data on the accuracy of load forecasting results, and select the fluctuation characteristics with the quantification value of the influence greater than a preset threshold from the fluctuation characteristics of the various types of data to obtain the core information feature set. The correlation module is used to perform time-series correlation processing on the core information feature set using the GRU time-series prediction model, to obtain the time-series correlation rules between different fluctuation features in the core information feature set, and to generate the load prediction results of the virtual power plant within a preset time period based on the time-series correlation rules.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the load accurate prediction method for virtual power plant multi-source heterogeneous data fusion as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the accurate load forecasting method for multi-source heterogeneous data fusion of a virtual power plant as described in any one of claims 1 to 6.

Images