Power grid intelligent load forecasting method based on deep learning and multi-source data fusion

Abstract

The present application relates to the technical field of power grid load prediction, and more particularly to a power grid intelligent load prediction method based on deep learning and multi-source data fusion, comprising: generating a plurality of key data sets based on the node characteristics of each power grid node; determining the data sensitivity corresponding to each key data set and determining the power intervention category corresponding to each key data set; monitoring the data of each power grid node to determine the power fluctuation risk index corresponding to each power grid node; determining a plurality of abnormal power grid nodes based on the power fluctuation risk index and determining a plurality of abnormal key data sets based on the data sensitivity corresponding to each abnormal power grid node and each key data set; constructing a strong intervention prediction model and a weak intervention prediction model to obtain the target predicted load in a future preset time period. The present application can improve the accuracy of power grid load prediction.

Description

Technical Field

[0001] This invention relates to the field of power grid load forecasting technology, and in particular to a smart power grid load forecasting method based on deep learning and multi-source data fusion. Background Technology

[0002] In power systems, industrial parks and other similar areas have numerous grid nodes, resulting in complex and rapidly changing electricity loads. Traditional load forecasting methods primarily rely on historical load data, lacking comprehensive consideration of multi-source data. This makes it difficult to accurately predict load changes, especially when facing issues such as equipment failures, grid overload, and inadequate power facility planning. These methods fail to detect anomalies promptly and make effective predictions, increasing the risk of power outages and causing significant losses to businesses and residents. Therefore, a more accurate and intelligent load forecasting method is needed, capable of comprehensively considering multiple factors, monitoring grid status in real time, and effectively identifying and predicting anomalies.

[0003] Chinese Patent Application Publication No. CN109816164A discloses a power load forecasting method, comprising: step S100, acquiring historical power load data for n historical periods of the current period, where Di is the power load data for the (n+1-i)th period before the current period, 1≤i≤n; step S200, acquiring historical high-frequency time-domain data and historical low-frequency time-domain data based on the historical power load data D using a spectrum analysis method; step S300, forecasting the power load data for the current period based on the historical high-frequency time-domain data DH and the historical low-frequency time-domain data DL, where p1 is the low-frequency predicted value obtained based on the historical low-frequency time-domain data DL, and phj is the high-frequency predicted value obtained based on the historical high-frequency time-domain data DHj.

[0004] The existing technology has the following problems: it only decomposes historical power load data based on spectrum analysis and makes predictions on the high-frequency and low-frequency parts separately to obtain the power load data for the current period. However, high-frequency time domain data usually contains a lot of noise and random fluctuations, and although low-frequency time domain data can reflect long-term trends, it has a certain lag and cannot capture short-term load changes in time, resulting in large prediction errors and low accuracy. Summary of the Invention

[0005] To address this issue, the present invention provides a smart load forecasting method for power grids based on deep learning and multi-source data fusion, which overcomes the problem in existing technologies that rely solely on spectrum analysis to decompose historical power load data and forecast high-frequency and low-frequency components separately, resulting in large forecasting errors and low accuracy.

[0006] To achieve the above objectives, this invention provides a smart load forecasting method for power grids based on deep learning and multi-source data fusion, comprising:

[0007] Step S1: Obtain historical power data of several power grid nodes within the target area, and generate several key datasets based on the node characteristics of each power grid node, wherein each key dataset includes historical power data of at least one power grid node.

[0008] Step S2: Determine the data sensitivity of each key dataset based on the historical power data in each key dataset, and determine the power interference category corresponding to each key dataset;

[0009] Step S3: Monitor the data of each power grid node and obtain the power data fluctuation curve of each power grid node within the target time period to determine the power fluctuation risk index of each power grid node.

[0010] Step S4: Based on each of the power fluctuation risk indices, determine a number of abnormal power grid nodes, and based on the data sensitivity of each abnormal power grid node and each key dataset, determine a number of abnormal key datasets.

[0011] Step S5: Based on the power interference category corresponding to each of the aforementioned key abnormal datasets, determine the strong interference dataset and the weak interference dataset respectively, so as to construct the strong interference prediction model and the weak interference prediction model.

[0012] Step S6: Input the real-time power data of each abnormal power grid node into the strong interference prediction model and the weak interference prediction model respectively to obtain the target predicted load within a preset time period in the future.

[0013] Furthermore, in step S1, generating several key datasets includes:

[0014] Step S11: Determine the node characteristics of each power grid node based on the historical power data of each power grid node;

[0015] Step S12: Cluster the node characteristics of each power grid node to generate several node cluster groups, wherein each node cluster group includes several power grid nodes.

[0016] Step S13: Generate several key datasets based on the historical power data of the power grid nodes in each of the node cluster groups.

[0017] Further, in step S2, determining the data sensitivity corresponding to each of the key datasets includes:

[0018] Step S21: Based on the historical power data of each power grid node in each of the key datasets, determine the degree of node influence corresponding to each power grid node in each key dataset.

[0019] Step S22: Based on the degree of node influence corresponding to each power grid node in each of the key datasets, determine the data sensitivity corresponding to each of the key datasets.

[0020] Further, step S2 includes:

[0021] Based on the comparison results between the data sensitivity corresponding to each key dataset and the first preset sensitivity, the power interference category corresponding to each key dataset is determined;

[0022] The electrical interference categories include strong interference and weak interference.

[0023] Further, in step S3, determining the power fluctuation risk index corresponding to each power grid node includes:

[0024] Step S31: Determine the time period of power anomaly for each power grid node based on the power data fluctuation curve corresponding to each power grid node.

[0025] Step S32: Determine the power fluctuation risk index corresponding to each power grid node based on the power anomaly time period corresponding to each power grid node.

[0026] Further, in step S4, determining a plurality of the abnormal power grid nodes includes:

[0027] Based on the comparison results between the power fluctuation risk index corresponding to each power grid node and the preset risk index, several abnormal power grid nodes are identified.

[0028] Furthermore, in step S4, the identification of several anomalous key datasets includes:

[0029] Step S41: Determine several abnormal datasets based on the key datasets corresponding to each of the abnormal power grid nodes.

[0030] Step S42: Determine several key abnormal datasets based on the comparison results between the data sensitivity corresponding to each of the abnormal datasets and the second preset sensitivity.

[0031] The first preset sensitivity is greater than the second preset sensitivity.

[0032] Further, step S5 includes:

[0033] If the power interference category corresponding to the abnormal key dataset is a strong interference category, then a strong interference dataset is determined based on the abnormal key dataset, and an initial deep learning model is trained based on the strong interference dataset to obtain a strong interference prediction model.

[0034] Further, step S5 includes:

[0035] If the power interference category corresponding to the abnormal key dataset is a weak interference category, then a weak interference dataset is determined based on the abnormal key dataset, and an initial deep learning model is trained based on the weak interference dataset to obtain a weak interference prediction model.

[0036] Further, step S6 includes:

[0037] Step S61: Input the real-time power data of each abnormal power grid node into the strong interference prediction model to obtain the strong interference prediction load output by the strong interference prediction model.

[0038] Step S62: Input the real-time power data of each abnormal power grid node into the weak interference prediction model to obtain the weak interference prediction load output by the weak interference prediction model.

[0039] Step S63: Based on the comparison results of the strong interference predicted load and the weak interference predicted load, determine the target predicted load within a preset time period in the future.

[0040] Compared with existing technologies, the advantages of this invention are as follows: By integrating multi-source data and generating several key datasets based on the node characteristics of each power grid node, this invention can more comprehensively reflect the characteristics of power grid nodes. Grouping power grid node data avoids data mixing based on different node characteristics, providing a structured data foundation for subsequent analysis and modeling. By assessing data sensitivity and determining the power interference category corresponding to each key dataset, the impact of data in different key datasets on power grid operation can be accurately located, providing a basis for subsequent model classification and avoiding insufficient adaptability of a single model. By monitoring data from each power grid node and calculating the power fluctuation risk index through power data fluctuation curves, the power grid status can be dynamically monitored, power fluctuations can be responded to in a timely manner, and the fluctuation risk of each node can be quantified. By screening abnormal power grid nodes using the power fluctuation risk index and then prioritizing highly sensitive abnormal key datasets based on data sensitivity, the accuracy of subsequent analysis and prediction is ensured, and interference from irrelevant data is avoided. By constructing strong interference prediction models and weak interference prediction models based on different categories of power interference, and modeling data for different interference categories separately, we can avoid errors caused by a single model's inability to capture sudden changes in strong interference scenarios, or overfitting due to excessive complexity in weak interference scenarios. The two prediction models can complement each other, and combining them for load forecasting can improve the accuracy of power grid load forecasting.

[0041] Furthermore, this invention extracts node features from historical power data to achieve data feature transformation, avoiding clustering bias caused by the high dimensionality and redundant information of the original time-series data. Through cluster analysis, power grid nodes with similar features are grouped, which can organize the data more effectively and ensure that power grid nodes of the same type have highly similar power consumption patterns and influencing factors. The resulting key dataset has less internal noise and higher feature correlation, providing a data foundation for subsequent data sensitivity analysis and interference category determination. A key dataset is generated for each cluster group, which facilitates the optimization of models for different load patterns and improves the adaptability and prediction accuracy of the models.

[0042] Furthermore, this invention analyzes historical power data to quantify the impact of each power grid node on the overall load, identifies nodes that have a significant impact on load changes, and assesses the overall sensitivity of each key dataset to load changes by comprehensively considering the impact of each power grid node in the key dataset, thereby optimizing resource allocation and improving the adaptability and accuracy of the prediction model.

[0043] Furthermore, by analyzing power data fluctuation curves, this invention can accurately identify the power anomaly periods of each power grid node, comprehensively analyze the key characteristics of the power anomaly periods to determine the power fluctuation risk index corresponding to each power grid node, quantify the impact risk of power anomaly periods, optimize resource allocation, and improve the accuracy of anomaly monitoring.

[0044] Furthermore, this invention avoids isolated analysis by associating abnormal power grid nodes with their corresponding key datasets, thereby uncovering the group correlations of anomalies and improving data processing efficiency. By comparing data sensitivity with a second preset sensitivity, it further filters out abnormal key datasets that are more sensitive to load changes, enhancing model adaptability and further improving the relevance and effectiveness of the datasets, thus improving the adaptability and accuracy of subsequent prediction models.

[0045] Furthermore, this invention uses a strong interference prediction model to model datasets with high sensitivity and high impact, which can more accurately predict load changes under strong interference conditions, capture significant features in load changes, and improve prediction accuracy. Conversely, it uses a weak interference prediction model specifically for datasets with low sensitivity and low impact, which can more comprehensively cover all aspects of load changes. By comparing the load predictions of strong and weak interference, the invention can comprehensively consider both significant and subtle features of load changes, thereby improving the comprehensiveness and accuracy of predictions. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the intelligent load forecasting method for power grids based on deep learning and multi-source data fusion, as described in an embodiment of the present invention.

[0047] Figure 2A flowchart illustrating the process of generating several key datasets for embodiments of the present invention;

[0048] Figure 3 This is a flowchart illustrating the process of determining the data sensitivity corresponding to each key dataset in an embodiment of the present invention.

[0049] Figure 4 A logic diagram for determining the type of electrical interference in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0051] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0052] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0053] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0054] Please see Figure 1 The diagram shown is a flowchart illustrating the intelligent load forecasting method for power grids based on deep learning and multi-source data fusion according to an embodiment of the present invention. The present invention provides an intelligent load forecasting method for power grids based on deep learning and multi-source data fusion, comprising:

[0055] Step S1: Obtain historical power data of several power grid nodes within the target area, and generate several key datasets based on the node characteristics of each power grid node, wherein each key dataset includes historical power data of at least one power grid node.

[0056] In practice, power data includes, but is not limited to, active power, reactive power, current, voltage, frequency, reactive power, total power supply, total power consumption, line current carrying capacity, and total load.

[0057] Please see Figure 2 The diagram illustrates the process of generating several key datasets according to an embodiment of the present invention. Specifically, in step S1, generating several key datasets includes:

[0058] Step S11: Determine the node characteristics of each power grid node based on the historical power data of each power grid node;

[0059] In implementation, historical power data for each power grid node is preprocessed, including data cleaning and data normalization. Data cleaning includes handling missing values ​​and outliers. Missing values ​​can be filled using methods such as "mean imputation" or "mode imputation". Outliers in numerical features (such as extreme load values ​​caused by sensor failures) can be identified using the "Raida criterion" (3σ principle) and replaced with the median of the data. Data normalization can use "Z-score standardization" (transforming features into a distribution with mean = 0 and standard deviation = 1) to eliminate the influence of dimensions. This is existing technology and will not be elaborated further.

[0060] Understandably, the preprocessed historical power data of each power grid node is used to perform feature selection through principal component analysis or correlation analysis to determine the node characteristics of each power grid node and integrate them into a node feature vector.

[0061] Step S12: Cluster the node characteristics of each power grid node to generate several node cluster groups, wherein each node cluster group includes several power grid nodes.

[0062] In practice, those skilled in the art will recognize that any existing clustering method capable of grouping power grid nodes with similar node characteristics into a single group falls within the scope of protection of this invention. Examples include K-Means and DBSCAN methods, which will not be elaborated upon here. Each node cluster group must include at least one power grid node.

[0063] Step S13: Generate several key datasets based on the historical power data of the power grid nodes in each of the node cluster groups.

[0064] In practice, the historical power data of the power grid nodes in any node cluster group constitute the key dataset corresponding to that node cluster group.

[0065] This invention extracts node features from historical power data, achieving data feature transformation and avoiding clustering bias caused by the high dimensionality and redundant information of the original time-series data. Through cluster analysis, power grid nodes with similar features are grouped, enabling more effective data organization and ensuring that similar power grid nodes have highly similar electricity consumption patterns and influencing factors. The resulting key dataset has less internal noise and higher feature correlation, providing a data foundation for subsequent data sensitivity analysis and interference category determination. A key dataset is generated for each cluster group, facilitating the optimization of models for different load patterns and improving the model's adaptability and prediction accuracy.

[0066] Step S2: Determine the data sensitivity of each key dataset based on the historical power data in each key dataset, and determine the power interference category corresponding to each key dataset;

[0067] Please see Figure 3 The diagram illustrates the process of determining the data sensitivity corresponding to each key dataset in an embodiment of the present invention. Specifically, in step S2, determining the data sensitivity corresponding to each key dataset includes:

[0068] Step S21: Based on the historical power data of each power grid node in each of the key datasets, determine the degree of node influence corresponding to each power grid node in each key dataset.

[0069] In one specific embodiment, a linear or nonlinear regression model is used, with the node characteristics of the power grid node as the independent variable and the total load as the dependent variable, to calculate the regression coefficient of each power grid node, which is then used as a quantitative indicator of the degree of node influence.

[0070] In another specific embodiment, the correlation between each node characteristic and the total load can be calculated using methods such as Pearson correlation coefficient and Spearman rank correlation coefficient. The correlation coefficient of each power grid node can be calculated and used as a quantitative indicator of the degree of node influence.

[0071] Step S22: Based on the degree of node influence corresponding to each power grid node in each of the key datasets, determine the data sensitivity corresponding to each of the key datasets.

[0072] In implementation, the average value of the node influence degree corresponding to each power grid node in any key dataset is determined as the data sensitivity corresponding to that key dataset.

[0073] This invention analyzes historical power data to quantify the impact of each power grid node on the overall load, identifies nodes that have a significant impact on load changes, and assesses the overall sensitivity of each key dataset to load changes by comprehensively considering the impact of each power grid node in the key dataset, thereby optimizing resource allocation and improving the adaptability and accuracy of the prediction model.

[0074] Please see Figure 4 As shown, it is a logic judgment diagram for determining the type of power interference in an embodiment of the present invention; specifically, step S2 includes:

[0075] Based on the comparison results between the data sensitivity corresponding to each key dataset and the first preset sensitivity, the power interference category corresponding to each key dataset is determined;

[0076] The electrical interference categories include strong interference and weak interference.

[0077] In implementation, for any key dataset, if the data sensitivity corresponding to the key dataset is greater than the first preset sensitivity, then the power interference category corresponding to the key dataset is determined as a strong interference category; if the data sensitivity corresponding to the key dataset is less than or equal to the first preset sensitivity, then the power interference category corresponding to the key dataset is determined as a strong interference category.

[0078] Understandably, implementers can set the first preset sensitivity based on the actual situation or the average sensitivity of data that passed the qualification test in historical data.

[0079] Step S3: Monitor the data of each power grid node and obtain the power data fluctuation curve of each power grid node within the target time period to determine the power fluctuation risk index of each power grid node.

[0080] Specifically, in step S3, determining the power fluctuation risk index corresponding to each power grid node includes:

[0081] Step S31: Determine the time period of power anomaly for each power grid node based on the power data fluctuation curve corresponding to each power grid node.

[0082] Step S32: Determine the power fluctuation risk index corresponding to each power grid node based on the power anomaly time period corresponding to each power grid node.

[0083] In implementation, the power data fluctuation curve uses time as the independent variable and power data as the dependent variable. For example, the active power fluctuation curve uses time as the independent variable and active power as the dependent variable. The power anomaly time period is the time period with the largest power data mutation. It can be determined based on the slope or standard deviation of the power data fluctuation curve. By setting a slope threshold or standard deviation threshold, the time period that exceeds the slope threshold (or standard deviation threshold) and contains the maximum slope (or maximum standard deviation) is determined as the power anomaly time period. Each grid node can have zero or one or more power anomaly time periods. If any grid node has no power anomaly time periods, the power fluctuation risk index corresponding to that grid node is determined to be 0. If any grid node has one or more power anomaly time periods, the ratio of the total duration of the power anomaly time periods (if there are multiple power anomaly time periods, the time periods are merged, and overlapping time periods are only calculated once) to the total duration of the target time period is determined as the power fluctuation risk index corresponding to that grid node.

[0084] This invention analyzes power data fluctuation curves to accurately identify abnormal power periods at each power grid node. By comprehensively analyzing the key characteristics of these abnormal power periods, it determines the power fluctuation risk index corresponding to each power grid node, quantifies the impact risk of abnormal power periods, optimizes resource allocation, and improves the accuracy of anomaly monitoring.

[0085] Step S4: Based on each of the power fluctuation risk indices, determine a number of abnormal power grid nodes, and based on the data sensitivity of each abnormal power grid node and each key dataset, determine a number of abnormal key datasets.

[0086] Specifically, in step S4, determining a plurality of the abnormal power grid nodes includes:

[0087] Based on the comparison results between the power fluctuation risk index corresponding to each power grid node and the preset risk index, several abnormal power grid nodes are identified.

[0088] In implementation, if the power fluctuation risk index corresponding to any power grid node is greater than the preset risk index, the power grid node is identified as an abnormal power grid node. The actual implementers can set the preset risk index based on the actual situation or the average value of the power fluctuation risk index that has passed the qualification test in historical data. Preferably, the preset risk index is set to a range of 0.6 to 0.7.

[0089] Specifically, in step S4, identifying several anomalous key datasets includes:

[0090] Step S41: Determine several abnormal datasets based on the key datasets corresponding to each of the abnormal power grid nodes.

[0091] Step S42: Determine several key abnormal datasets based on the comparison results between the data sensitivity corresponding to each of the abnormal datasets and the second preset sensitivity.

[0092] The first preset sensitivity is greater than the second preset sensitivity.

[0093] In implementation, the key dataset corresponding to any abnormal power grid node is identified as the abnormal dataset (i.e., the abnormal dataset is a key dataset that includes historical power data of the abnormal power grid node).

[0094] It is understandable that if the data sensitivity corresponding to any abnormal dataset is greater than the second preset sensitivity, then the abnormal dataset is identified as an abnormal critical dataset. The implementers can set the second preset sensitivity based on the actual situation, or set the second preset sensitivity to 2 / 3 to 3 / 4 of the first preset sensitivity.

[0095] This invention avoids isolated analysis by associating abnormal power grid nodes with their corresponding key datasets, thereby uncovering the group correlations of anomalies and improving data processing efficiency. By comparing data sensitivity with a second preset sensitivity, it further filters out abnormal key datasets that are more sensitive to load changes, enhancing model adaptability and further improving the relevance and effectiveness of the datasets, thus improving the adaptability and accuracy of subsequent prediction models.

[0096] Step S5: Based on the power interference category corresponding to each of the aforementioned key abnormal datasets, determine the strong interference dataset and the weak interference dataset respectively, so as to construct the strong interference prediction model and the weak interference prediction model.

[0097] Specifically, step S5 includes:

[0098] If the power interference category corresponding to the abnormal key dataset is a strong interference category, then a strong interference dataset is determined based on the abnormal key dataset, and an initial deep learning model is trained based on the strong interference dataset to obtain a strong interference prediction model.

[0099] In practice, if the number of anomalous key datasets with the power interference category as strong interference is unique, then the anomalous key dataset is identified as a strong interference dataset. If the number of anomalous key datasets with the power interference category as strong interference is not unique, then the anomalous key datasets with the power interference category as strong interference are merged to obtain a strong interference dataset.

[0100] It should be noted that those skilled in the art will understand that any deep learning model in the prior art that can predict the load falls within the protection scope of this invention, such as the CNN-LSTM-Attention model, the Transformer model, the multilayer perceptron, etc., which will not be described in detail here.

[0101] Specifically, step S5 includes:

[0102] If the power interference category corresponding to the abnormal key dataset is a weak interference category, then a weak interference dataset is determined based on the abnormal key dataset, and an initial deep learning model is trained based on the weak interference dataset to obtain a weak interference prediction model.

[0103] In practice, if the number of anomalous key datasets whose power interference category is weak interference is unique, then the anomalous key dataset is identified as a weak interference dataset. If the number of anomalous key datasets whose power interference category is weak interference is not unique, then the anomalous key datasets whose power interference category is weak interference are merged to obtain a weak interference dataset.

[0104] It is understood that those skilled in the art will recognize that any existing deep learning model capable of predicting grid load falls within the protection scope of this invention, and will not be elaborated further here.

[0105] Step S6: Input the real-time power data of each abnormal power grid node into the strong interference prediction model and the weak interference prediction model respectively to obtain the target predicted load within a preset time period in the future.

[0106] Specifically, step S6 includes:

[0107] Step S61: Input the real-time power data of each abnormal power grid node into the strong interference prediction model to obtain the strong interference prediction load output by the strong interference prediction model.

[0108] Step S62: Input the real-time power data of each abnormal power grid node into the weak interference prediction model to obtain the weak interference prediction load output by the weak interference prediction model.

[0109] Step S63: Based on the comparison results of the strong interference predicted load and the weak interference predicted load, determine the target predicted load within a preset time period in the future.

[0110] In implementation, the absolute value of the difference between the strong interference prediction load and the weak interference prediction load is determined as the first difference, the mean of the strong interference prediction load and the weak interference prediction load is determined as the first mean, and the ratio of the first difference to the first mean is determined as the comprehensive evaluation index.

[0111] Understandably, if the comprehensive evaluation index is less than the preset evaluation index, it indicates that there is no significant change in the power data within the target time period, and the first average value is determined as the target predicted load. If the comprehensive evaluation index is greater than or equal to the preset evaluation index, it indicates that there may be significant changes in the power data within the target time period, and the strong interference predicted load is determined as the target predicted load.

[0112] Understandably, implementers can set preset evaluation indices based on actual conditions or the average comprehensive evaluation index of qualified products from historical data.

[0113] It is understood that the end time of the target time period is the current time point, the power data obtained at the current time point is real-time power data, and the future preset time period starts from the current time point. Preferably, the target time period is 30 min to 60 min, and the future preset time period is 10 min to 15 min.

[0114] This invention employs a strong interferometry prediction model to model high-sensitivity and high-impact datasets, enabling more accurate prediction of load changes under strong interferometry conditions, capturing significant features in load changes, and improving prediction accuracy. Conversely, a weak interferometry prediction model is specifically designed for low-sensitivity and low-impact datasets, providing more comprehensive coverage of all aspects of load changes. By comparing the load predictions from strong and weak interferometry, the invention comprehensively considers both significant and subtle features of load changes, improving the comprehensiveness and accuracy of predictions.

[0115] Specifically, this invention integrates multi-source data and generates several key datasets based on the node characteristics of each power grid node. This more comprehensively reflects the characteristics of power grid nodes, groups the power grid node data to avoid data mixing of different node characteristics, and provides a structured data foundation for subsequent analysis and modeling. By assessing data sensitivity and determining the power interference category corresponding to each key dataset, the impact of data in different key datasets on power grid operation can be accurately located, providing a basis for subsequent model classification and avoiding the insufficient adaptability of a single model. By monitoring data from each power grid node and calculating the power fluctuation risk index through power data fluctuation curves, the power grid status can be dynamically monitored, power fluctuations can be responded to in a timely manner, and the fluctuation risk of each node can be quantified. By screening abnormal power grid nodes through the power fluctuation risk index and then prioritizing the identification of highly sensitive abnormal key datasets based on data sensitivity, the accuracy of subsequent analysis and prediction is ensured, and interference from irrelevant data is avoided. By constructing strong interference prediction models and weak interference prediction models based on different categories of power interference, and modeling data for different interference categories separately, we can avoid errors caused by a single model's inability to capture sudden changes in strong interference scenarios, or overfitting due to excessive complexity in weak interference scenarios. The two prediction models can complement each other, and combining them for load forecasting can improve the accuracy of power grid load forecasting.

[0116] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A power grid intelligent load prediction method based on deep learning and multi-source data fusion, characterized in that, The method comprises the following steps: Step S1, obtaining historical power data of a plurality of power grid nodes in a target area, and generating a plurality of key data sets based on the node characteristics of each power grid node, wherein each key data set comprises historical power data of at least one power grid node; Step S2, determining the data sensitivity corresponding to each key data set based on the historical power data in each key data set, and determining the power interference category corresponding to each key data set; Step S3, respectively monitoring the data of each power grid node, obtaining the power data fluctuation curve of each power grid node in a target time period, and determining the power fluctuation risk index corresponding to each power grid node; Step S4, determining a plurality of abnormal power grid nodes based on the power fluctuation risk index, and determining a plurality of abnormal key data sets based on the abnormal power grid nodes and the data sensitivity corresponding to each key data set; Step S5, respectively determining strong interference data sets and weak interference data sets based on the power interference category corresponding to each abnormal key data set, to construct a strong interference prediction model and a weak interference prediction model; Step S6, inputting the real-time power data of each abnormal power grid node into the strong interference prediction model and the weak interference prediction model respectively, to obtain the target predicted load in a future preset time period; In the step S2, it comprises: Based on the comparison result of the data sensitivity corresponding to each key data set and the first preset sensitivity, the power interference category corresponding to each key data set is determined; Wherein, the power interference category includes strong interference category and weak interference category; In the step S3, determining the power fluctuation risk index corresponding to each power grid node comprises: Step S31, determining the power anomaly time period corresponding to each power grid node based on the power data fluctuation curve corresponding to each power grid node; Step S32, determining the power fluctuation risk index corresponding to each power grid node based on the power anomaly time period corresponding to each power grid node; In the step S4, determining a plurality of abnormal key data sets comprises: Step S41, determining a plurality of abnormal data sets based on the key data sets corresponding to each abnormal power grid node; Step S42, determining a plurality of abnormal key data sets based on the comparison result of the data sensitivity corresponding to each abnormal data set and the second preset sensitivity; Wherein, the first preset sensitivity is greater than the second preset sensitivity.

2. The power grid intelligent load forecasting method based on deep learning and multi-source data fusion according to claim 1, characterized in that, In the step S1, generating a plurality of key data sets comprises: Step S11, determining the node characteristics of each power grid node based on the historical power data of each power grid node; Step S12, clustering the node characteristics of each power grid node to generate a plurality of node clustering groups, wherein each node clustering group comprises a plurality of power grid nodes; Step S13, generating a plurality of key data sets based on the historical power data of the power grid nodes in each node clustering group.

3. The power grid intelligent load forecasting method based on deep learning and multi-source data fusion according to claim 2, characterized in that, In the step S2, determining the data sensitivity corresponding to each key data set comprises: Step S21, respectively determining the node influence degree corresponding to each power grid node in each key data set based on the historical power data of each power grid node in each key data set; In step S22, data sensitivity corresponding to each of the key data sets is determined based on the node influence degree corresponding to each power grid node in each of the key data sets.

4. The power grid intelligent load forecasting method based on deep learning and multi-source data fusion according to claim 3, characterized in that, In the step S4, the determination of the abnormal power grid nodes includes: Based on the comparison result of the power fluctuation risk index corresponding to each of the power grid nodes and the preset risk index, the abnormal power grid nodes are determined.

5. The power grid intelligent load forecasting method based on deep learning and multi-source data fusion according to claim 4, characterized in that, In the step S5, it includes: If the power intervention category corresponding to the abnormal key data set is the strong intervention category, a strong intervention data set is determined based on the abnormal key data set, and an initial deep learning model is trained based on the strong intervention data set to obtain a strong intervention prediction model.

6. The power grid intelligent load forecasting method based on deep learning and multi-source data fusion according to claim 5, characterized in that, In the step S5, it includes: If the power intervention category corresponding to the abnormal key data set is the weak intervention category, a weak intervention data set is determined based on the abnormal key data set, and an initial deep learning model is trained based on the weak intervention data set to obtain a weak intervention prediction model.

7. The power grid intelligent load forecasting method based on deep learning and multi-source data fusion according to claim 6, characterized in that, In the step S6, it includes: In step S61, real-time power data of each of the abnormal power grid nodes is input into the strong intervention prediction model to obtain a strong intervention prediction load output by the strong intervention prediction model; In step S62, real-time power data of each of the abnormal power grid nodes is input into the weak intervention prediction model to obtain a weak intervention prediction load output by the weak intervention prediction model; In step S63, based on the comparison result of the strong intervention prediction load and the weak intervention prediction load, a target prediction load in a future preset time period is determined.

Images