Power system load prediction method based on multivariate data fusion

Abstract

The invention discloses a power system load prediction method based on multivariate data fusion, and relates to the technical field of power systems. The problems that in the prior art, multi-source data fusion processing is complex, timeliness and processing efficiency are insufficient, and under the complex working condition, the response capacity to short-term emergencies is weak are solved. By collecting and preprocessing the multi-source data in real time, various factors influencing load changes can be comprehensively considered, the response capability to complex working conditions and emergencies is enhanced, the time sequence window is dynamically adjusted to improve the data validity and prediction performance, the accuracy and reliability of power system load prediction are improved, and the power system load prediction efficiency is improved. According to the method, multiple models are constructed and trained to obtain an accurate load prediction result, intelligent prediction and management of the load of the power system are realized through accurate load prediction, abnormal identification and processing are performed in combination with an emergency database, stable and reliable load prediction is provided for the power system, and thus power resource configuration is optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a power system load forecasting method based on multivariate data fusion. Background Art

[0002] Traditional load forecasting methods often rely only on historical load data, ignoring other factors that may affect load changes, such as weather conditions, socioeconomic conditions, user behavior patterns, etc. The Chinese patent with publication number: CN118780449A discloses a method for power load forecasting, including obtaining power system data; preprocessing the power system data and multi-source data fusion processing; using the variational mode decomposition model to decompose the original time series into subsequences; constructing a power load forecasting model based on the gated recurrent unit and the attention mechanism, and using the power load forecasting model to predict the power load based on the decomposed subsequences, wherein the parameters of the power load forecasting model are optimized using an improved sparrow search algorithm. This patent solves the problems of insufficient accuracy of traditional power load forecasting models, difficulty in optimizing model parameters, and poor multi-source data fusion processing effects.

[0003] However, although the above patents have improved the prediction accuracy to a certain extent, they still have the following problems: The multi-source data fusion processing in the existing technology is complex, its timeliness and processing efficiency are insufficient, and under complex working conditions, its response capability to short-term emergencies is weak. Summary of the invention

[0004] The purpose of the present invention is to provide a power system load forecasting method based on multivariate data fusion. By integrating and analyzing information from multiple data sources, it can comprehensively consider various factors affecting load changes, enhance the response capability to complex working conditions and emergencies, and improve the accuracy and reliability of power system load forecasting. Through accurate load forecasting, intelligent prediction and management of power system load is realized to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: The power system load forecasting method based on multivariate data fusion includes the following steps: Step 1: Data preprocessing: Collect multi-source data in real time based on the multi-source acquisition terminal, clean and normalize the acquired multi-source data, and perform feature coding based on the multi-source data category; Step 2: Multi-source data fusion: Fusion and feature optimization of pre-processed multi-source data, screening out key features that are highly relevant to load forecasting, and dynamically adjusting the timing window based on short-term and long-term load forecasting requirements; Step 3: Multi-model load forecasting: Build and train the corresponding load forecasting model based on multi-source data categories, and use the trained load forecasting model to perform load forecasting to obtain the forecast results of future loads; Step 4: Forecast result evaluation: Evaluate the model accuracy through model evaluation indicators, adjust the structure and parameters of the load forecast model based on the evaluation results and combined with real-time forecast accuracy feedback data, and visualize the final future load forecast results; Step 5: Abnormal response: Establish an emergency event database, identify the abnormal type of the prediction result based on the emergency event database, and perform abnormal processing response according to the abnormal type.

[0006] Furthermore, the real-time collection of multi-source data based on the multi-source collection terminal specifically includes: Determine the data source: collect historical load data, meteorological data, holiday information and economic indicator data of the power grid in real time; Data preprocessing: remove duplicates and irrelevant information from the collected multi-source data, and standardize and convert multi-source data of different dimensions; Feature encoding processing: Perform one-hot encoding and label encoding processing on the category features of multi-source data. Among them, one-hot encoding is used to convert the category features of holiday information and meteorological data into binary vectors, and label encoding is used to process the categorical variables of historical load data; Consistency check: Align the timestamps and sampling frequencies carried by multi-source data, check and unify the measurement units and scales of multi-source data, and generate continuous and synchronized time series data.

[0007] Further, step 2: multi-source data fusion, specifically: Multi-source data fusion: After consistency verification, multi-source data is fused at the feature level to build a unified feature matrix; Feature identification: Calculate the correlation coefficient between each feature in the feature matrix and the power load, and identify the key features related to the power load; Feature screening: Calculate the importance weight of each key feature and recursively remove the key features with the lowest weight until the preset number of features is reached to determine the final key feature subset; Dynamic adjustment of time window: Dynamically adjust the time window size according to the load forecasting requirements of different time granularities, determine the changing characteristics of load data at different time lengths, and use the sliding window method to optimize the window length based on the forecast accuracy feedback data.

[0008] Furthermore, the timing window is dynamically adjusted, including: Extract the change rate corresponding to the current load data; Comparing the change rate corresponding to the load data with a preset change rate threshold; When the change rate corresponding to the load data exceeds a preset change rate threshold, the time granularity corresponding to the current time series window is retrieved; Taking the time granularity corresponding to the duration of the current timing window as the unit reference time; Extract load data in each unit reference time; Obtaining a load factor corresponding to each unit reference time by using the load data in each unit reference time; The load factor corresponding to each unit reference time is obtained by the following formula:

[0009] Among them, U represents the load factor corresponding to each unit reference time; n represents the number of load data collections corresponding to each unit reference time; P represents the corresponding change rate value when exceeding the preset change rate threshold; W i represents the load data value corresponding to the i-th load data collection; W i+1 represents the load data value corresponding to the i+1th load data collection; W z represents the middle value of the load data corresponding to n load data collections; P z It represents the change rate between the load data corresponding to each unit reference time and the load data corresponding to the previous unit reference time; Retrieve the middle value of the load factor from the load factor corresponding to each unit reference time; comparing the load factor intermediate value with a preset load factor threshold; When the load factor intermediate value exceeds a preset load factor threshold, the timing window is dynamically adjusted.

[0010] Furthermore, the timing window is dynamically adjusted, including: When the load factor intermediate value exceeds the preset load factor threshold, the maximum time granularity value and the minimum time granularity value corresponding to the historical load data are retrieved; Taking the duration corresponding to the maximum value of the time granularity as the first unit reference time; Taking the duration corresponding to the minimum value of the time granularity as the second unit reference time; Under the first unit reference time, obtaining a load factor corresponding to each first unit reference time as a first load factor data set; Under the second unit reference time, obtaining a load factor corresponding to each second unit reference time as a second load factor data set; Obtaining a time window adjustment coefficient using the load factors included in the first load factor data set and the second load factor data set combined with an intermediate value of the load factors corresponding to the current unit reference time; The time window adjustment coefficient is obtained by the following formula:

[0011] Wherein, K represents the time window adjustment coefficient; m represents the number of load factors included in the first load factor data set; k represents the number of load factors included in the second load factor data set; U z Indicates the middle value of the load factor corresponding to the current unit reference time; U 01i represents the value corresponding to the i-th load factor in the first load factor data set; U 02i represents the value corresponding to the i-th load factor in the second load factor data set; U 01z represents the middle value of the coefficient corresponding to the m load factors in the first load factor data set; U 02z represents the middle value of the coefficient corresponding to the k load factors in the second load factor data set; T 01 and T 02 Respectively represent the corresponding durations of the first unit reference time and the second unit reference time; Using the time window adjustment coefficient to adjust the time granularity of the time window, and obtain the adjusted time granularity; The corresponding duration of the adjusted time granularity is obtained by the following formula:

[0012] Among them, T t Indicates the duration corresponding to the time granularity after adjustment; T0 indicates the duration corresponding to the time granularity before adjustment; K indicates the time window adjustment coefficient.

[0013] Furthermore, the feature recognition further includes: Acquire power data from the power system, perform redundancy and dimensionality reduction processing on the power data, perform feature extraction on the processed power data, and obtain an initial feature set of the power data according to the extraction result; Identify and extract key features from the initial feature set, and use each key feature in the key feature subset as a model output sample to train a preset network model to generate a load recognition model; Obtaining target power consumption characteristics corresponding to target time series data of each key feature in the key feature subset based on the load identification model; The change of the target power consumption characteristic of each load in the power data is obtained, and the power consumption change rule of each load is determined according to the change.

[0014] Furthermore, the prediction result of the future load in step 3 includes: Obtain the prediction accuracy of each key feature as the first indicator for evaluating model performance; Calculate the dynamic change characteristic value of each key feature as the second indicator for model performance evaluation; Count the frequency and data coverage of each key feature in the data set as a supplementary performance indicator; According to the prediction accuracy, dynamic characteristic value and data coverage of each key feature, the confidence of the corresponding feature is calculated and converted into a weight value; Based on the weight value, the prediction results of the power data corresponding to each data tag are weighted averaged to obtain the final load forecast result, and the importance of each key feature in the forecast result is quantified; Based on the weight value of each key feature, the prediction results of each key feature output by the load forecasting model are weighted averaged, and the final future load forecast result is generated based on the weighted calculation result.

[0015] Furthermore, the step 3 generates the final future load forecast result, and also includes determining the change pattern of the power load and identifying the power consumption pattern. The specific steps are as follows: Analysis of electricity consumption change patterns: Based on the electricity consumption variation rules of each load, identify and classify the electricity consumption characteristics during peak, valley and off-peak periods, and analyze the load fluctuation patterns during different periods; Combine meteorological data to determine the changing pattern of load affected by meteorological factors, identify the causes of seasonal peaks and troughs, and combine holiday information; Analyze the load fluctuation characteristics during holidays based on holiday information, distinguish the power consumption patterns during normal working days and holidays, and identify the load change patterns during special periods; Economic activity correlation analysis: Based on the correlation between economic indicator data evaluation and load power consumption change rules, a correlation model between load change and economic indicator data is established to predict the long-term growth trend of load; Comprehensive correlation analysis: Based on the fused multidimensional data, a multidimensional correlation model of power load changes is constructed to obtain the coupling relationship between load changes and multidimensional data; The identified electricity consumption variation patterns are verified based on the trained load forecasting model, and the electricity consumption pattern is optimized by combining multi-source data obtained in real time.

[0016] Furthermore, the model evaluation indicators in step 4 specifically include: mean absolute error indicator, mean square error indicator, root mean square error indicator, and mean absolute percentage error indicator.

[0017] Furthermore, the step 5 of identifying the abnormal type of the prediction result specifically includes: Classify and mark the emergency event information in the emergency event database, and associate the mark with the timestamp, impact area and load change data of the corresponding emergency event information; Compare the residuals between the actual load data and the predicted results, identify abnormal data points that exceed the normal fluctuation range, match the detected abnormal data points with the emergency event information, and identify the abnormal type based on the matching results; Re-evaluate the prediction results after exception handling, identify the change pattern of load before and after the event, and verify the effect of exception handling; A feedback mechanism between exception handling and load forecasting results is established, and the load forecasting model is adjusted based on real-time forecast accuracy feedback data.

[0018] Compared with the prior art, the present invention has the following beneficial effects: By collecting and preprocessing multi-source data in real time, it is possible to comprehensively consider various factors that affect load changes, enhance the ability to respond to complex working conditions and emergencies, dynamically adjust the timing window to improve data validity and prediction performance, and improve the accuracy and reliability of power system load forecasting. It builds and trains multiple models to obtain accurate load forecasting results, and uses evaluation indicators such as mean absolute error and mean square error to ensure prediction accuracy. Through accurate load forecasting, it realizes intelligent prediction and management of power system load, and combines with the emergency database for anomaly identification and processing, providing the power system with stable and reliable load forecasting, thereby optimizing the allocation of power resources and ensuring safe and economical operation of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The present invention is a flow chart of a method for power system load forecasting based on multivariate data fusion. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] In order to solve the technical problems of the existing technology in which multi-source data fusion processing is complex, its timeliness and processing efficiency are insufficient, and its ability to respond to short-term emergencies under complex working conditions is weak, please refer to Figure 1 , this embodiment provides the following technical solutions: The power system load forecasting method based on multivariate data fusion includes the following steps: Step 1: Data preprocessing: Based on the multi-source collection terminal, historical load data, meteorological data, holiday information, economic indicators and other multi-source data are collected in real time, and the acquired multi-source data are cleaned and normalized, missing values ​​and outliers are processed, and feature coding is performed based on multi-source data categories to ensure the consistency and integrity of multi-source data; Step 2: Multi-source data fusion: Fusion and feature optimization of pre-processed multi-source data, screening out key features that are highly relevant to load forecasting, and dynamically adjusting the timing window based on short-term and long-term load forecasting requirements to improve the effectiveness and forecasting performance of multi-source data; Step 3: Multi-model load forecasting: Build and train the corresponding load forecasting model based on the multi-source data categories, divide the preprocessed multi-source data into a training set and a test set, train the corresponding load forecasting model based on the training set data, and use the trained load forecasting model to perform load forecasting to obtain the prediction results of future loads; Among them, the load forecasting model includes time series model, regression model, neural network model and integrated learning model, etc. The training process includes data segmentation, loss function optimization, gradient descent and other processing, and adjusts the structure and parameters of the model through real-time feedback data.

[0022] In this embodiment, the time series model processes the time series data of the power load and can capture the seasonal and trend changes of the load, such as ARIMA, SARIMA, LSTM, etc., especially LSTM has a strong advantage in capturing long-term dependencies and dynamic change trends; In this embodiment, the regression model is used to establish the relationship between the load data and the multi-source features, such as linear regression, support vector regression SVR, XGBoost, etc.; In this embodiment, the neural network model automatically extracts complex features from multi-source data and optimizes prediction performance through a deep learning model, such as a multi-layer perceptron, a convolutional neural network (CNN), etc. In this embodiment, the ensemble learning model can improve the accuracy and stability of prediction and is suitable for the fusion of multiple features and complex pattern recognition, such as random forest, AdaBoost, etc.; In this embodiment, multiple models such as long short-term memory network (LSTM), support vector regression (SVR), random forest, etc. are used for load forecasting, so that it can process multi-source heterogeneous data and select and adjust the model according to the actual data characteristics and prediction accuracy.

[0023] Step 4: Forecast result evaluation: Evaluate the model accuracy through model evaluation indicators. Based on the evaluation results, adjust the structure and parameters of the load forecast model in combination with real-time forecast accuracy feedback data to ensure the accuracy and reliability of the forecast results, and visualize the final future load forecast results; The model evaluation indicators specifically include: mean absolute error indicator, which is used to measure the average deviation between the predicted value and the actual value; mean square error indicator, which is used to evaluate the average of the square of the difference between the predicted value and the actual value; root mean square error indicator, which indicates the size of the prediction error, and the dimension is consistent with the actual value; mean absolute percentage error indicator, which is used to measure the prediction accuracy; Step 5: Abnormal response: Establish an emergency event database to record events that may cause load abnormalities, such as weather disasters, equipment failures, etc., identify the abnormal type of the prediction results based on the emergency event database, and perform abnormal processing responses based on the abnormal type.

[0024] In this embodiment, by real-time collection and preprocessing of multi-source data, it is possible to comprehensively consider various factors affecting load changes, enhance the response capability to complex working conditions and emergencies, dynamically adjust the timing window to improve data validity and prediction performance, improve the accuracy and reliability of power system load forecasting, build and train multiple models to obtain accurate load forecasting results, and use evaluation indicators such as mean absolute error and mean square error to ensure prediction accuracy. Through accurate load forecasting, intelligent prediction and management of power system load is achieved, and abnormality identification and processing are performed in combination with the emergency event database, providing the power system with stable and reliable load forecasting, thereby optimizing the allocation of power resources and ensuring safe and economical operation of the system.

[0025] In this embodiment, the real-time collection of multi-source data based on the multi-source collection terminal specifically includes: Determine the data source: Collect the historical load data of the power grid in real time through the power system load monitoring terminal, obtain meteorological data such as temperature, humidity, wind speed, rainfall, etc. through meteorological stations, satellite remote sensing and other means, obtain holiday information from channels such as national statutory holidays and local special holidays, and collect economic indicator data such as GDP, industrial added value, and residents' consumption level released by the National Bureau of Statistics and local government departments; Data preprocessing: remove duplicates and irrelevant information from the collected multi-source data to ensure data quality, and standardize and convert multi-source data of different dimensions to facilitate subsequent analysis; Feature encoding processing: Perform one-hot encoding and label encoding processing on the category features of multi-source data. Among them, one-hot encoding is used to convert the category features of holiday information and meteorological data into binary vectors. Label encoding is used to process the categorical variables of historical load data, decompose the timestamps of historical load data, and extract time features, such as peak time, mid-peak time, and off-peak time. Consistency check: Align the timestamps and sampling frequencies carried by multi-source data to ensure the consistency of multi-source data in the time dimension, check and unify the measurement units and scales of multi-source data to ensure that the data has the same measurement standards, generate continuous and synchronized time series data, and use linear interpolation or forward filling methods for processing.

[0026] In this embodiment, by accurately determining the data source and collecting multi-source data in real time, efficient data preprocessing and quality assurance, feature coding processing and time feature extraction, as well as consistency verification and time series data generation, deep integration and efficient utilization of multi-source data are achieved, which significantly improves the accuracy and timeliness of power system load forecasting. It not only ensures the consistency and integrity of the data, but also enhances the model's ability to recognize complex load change patterns through refined feature processing, thereby effectively improving the accuracy of load forecasting, providing reliable data support and technical guarantee for the stable operation of the power system and optimal resource allocation, and has obvious practical application benefits.

[0027] In this embodiment, step 2: multi-source data fusion, specifically: Multi-source data fusion: After consistency check, multi-source data is fused at the feature level to build a unified feature matrix. The feature matrix includes all relevant features to ensure that all features are consistent in the same time scale, such as topological structure information, node attributes, and timing characteristics of the distribution network; Feature identification: Calculate the correlation coefficient between each feature in the feature matrix and the power load, and identify the key features related to the power load; Feature screening: Calculate the importance weight of each key feature, quantify the contribution of each feature to the prediction result, and recursively remove the key features with the lowest weight until the preset number of features is reached to determine the final key feature subset; Dynamic adjustment of time series window: Dynamically adjust the time window size according to the load forecasting requirements of different time granularities (such as hours, days, and weeks), determine the changing characteristics of load data at different time lengths, and use the sliding window method to optimize the window length according to the forecast accuracy feedback data to adapt to short-term and long-term load forecasting requirements and ensure that the time series characteristics are highly matched with load changes.

[0028] Specifically, the timing window is dynamically adjusted, including: Extract the change rate corresponding to the current load data; Comparing the change rate corresponding to the load data with a preset change rate threshold; When the change rate corresponding to the load data exceeds a preset change rate threshold, the time granularity corresponding to the current time series window is retrieved; Taking the time granularity corresponding to the duration of the current timing window as the unit reference time; Extract load data in each unit reference time; Obtaining a load factor corresponding to each unit reference time by using the load data in each unit reference time; The load factor corresponding to each unit reference time is obtained by the following formula:

[0029] Among them, U represents the load factor corresponding to each unit reference time; n represents the number of load data collections corresponding to each unit reference time; P represents the corresponding change rate value when exceeding the preset change rate threshold; W i represents the load data value corresponding to the i-th load data collection; W i+1 represents the load data value corresponding to the i+1th load data collection; W z represents the middle value of the load data corresponding to n load data collections; P z It represents the change rate between the load data corresponding to each unit reference time and the load data corresponding to the previous unit reference time; Retrieve the middle value of the load factor from the load factor corresponding to each unit reference time; comparing the load factor intermediate value with a preset load factor threshold; When the load factor intermediate value exceeds a preset load factor threshold, the timing window is dynamically adjusted.

[0030] The technical effect of the above technical solution is: by extracting the change rate of the current load data in real time and comparing it with the preset change rate threshold, the solution can dynamically identify the fluctuation of the load data. When the load data changes significantly (that is, the change rate exceeds the threshold), the solution can respond quickly and dynamically adjust the timing window. This dynamic adjustment mechanism enables the solution to adapt to different load change modes more flexibly, improving the adaptability and response speed of the system. The solution calculates the load factor in each unit reference time and uses the coefficient to evaluate the stability and change trend of the load data. The calculation of the load factor takes into account multiple factors such as the number of load data collection, the change rate value, and the intermediate value of the load data, so that the actual situation of the load data can be more accurately reflected. This refined processing method helps to improve the accuracy and accuracy of the system's analysis of load data. By taking the time granularity of the current timing window as the unit reference time and extracting the load data in each unit reference time, the solution can more effectively utilize data resources and reduce the interference of redundant information. At the same time, by comparing the intermediate value of the load factor with the preset load factor threshold, the timing window that needs to be adjusted can be further screened out, thereby improving the processing efficiency and stability of the system. The parameters in this solution, such as the rate of change threshold and load factor threshold, can be set and adjusted according to actual needs, making the solution scalable and customizable. This helps meet the needs of different application scenarios and improves the applicability and flexibility of the solution. By dynamically adjusting the timing window, the solution can allocate system resources more reasonably and avoid unnecessary calculations and processing when the load data fluctuates slightly. This helps reduce the energy consumption and cost of the system and improve resource utilization efficiency.

[0031] In summary, the technical effects of this technical solution in terms of performance indicators are mainly reflected in dynamic adaptability, precision and accuracy, efficiency and stability, scalability and customizability, as well as optimizing resource allocation and reducing energy consumption. These effects jointly improve the overall performance of the system and user experience.

[0032] Specifically, the timing window is dynamically adjusted, including: When the load factor intermediate value exceeds the preset load factor threshold, the maximum time granularity value and the minimum time granularity value corresponding to the historical load data are retrieved; Taking the duration corresponding to the maximum value of the time granularity as the first unit reference time; Taking the duration corresponding to the minimum value of the time granularity as the second unit reference time; Under the first unit reference time, obtaining a load factor corresponding to each first unit reference time as a first load factor data set; Under the second unit reference time, obtaining a load factor corresponding to each second unit reference time as a second load factor data set; Obtaining a time window adjustment coefficient using the load factors included in the first load factor data set and the second load factor data set combined with an intermediate value of the load factors corresponding to the current unit reference time; The time window adjustment coefficient is obtained by the following formula:

[0033] Wherein, K represents the time window adjustment coefficient; m represents the number of load factors included in the first load factor data set; k represents the number of load factors included in the second load factor data set; U z Indicates the middle value of the load factor corresponding to the current unit reference time; U 01i represents the value corresponding to the i-th load factor in the first load factor data set; U 02i represents the value corresponding to the i-th load factor in the second load factor data set; U 01z represents the middle value of the coefficient corresponding to the m load factors in the first load factor data set; U 02z represents the middle value of the coefficient corresponding to the k load factors in the second load factor data set; T 01 and T 02 Respectively represent the corresponding durations of the first unit reference time and the second unit reference time; Using the time window adjustment coefficient to adjust the time granularity of the time window, and obtain the adjusted time granularity; The corresponding duration of the adjusted time granularity is obtained by the following formula:

[0034] Among them, T t Indicates the duration corresponding to the time granularity after adjustment; T0 indicates the duration corresponding to the time granularity before adjustment; K indicates the time window adjustment coefficient.

[0035] The technical effect of the above technical solution is: the above technical solution realizes the dynamic adjustment of the time granularity of the time window by introducing the time window adjustment coefficient. When the intermediate value of the load factor exceeds the preset threshold, the above technical solution can automatically retrieve the time granularity information of the historical load data, and calculate the time window adjustment coefficient based on this information and the current intermediate value of the load factor. This dynamic adjustment mechanism enables the solution to better adapt to the fluctuation of load data and improve the accuracy and efficiency of analysis. By obtaining the load factor data set at the first unit reference time and the second unit reference time respectively, and combining the current intermediate value of the load factor, the solution can more comprehensively consider the changes of load data at different time granularities. This processing method helps to improve the accuracy and stability of load factor calculation, thereby more accurately reflecting the actual characteristics of load data. The above technical solution can reasonably allocate system resources according to the fluctuation of actual load data by dynamically adjusting the time granularity of the time window. When the load data fluctuates greatly, a smaller time granularity is used to capture data changes more finely; when the load data is relatively stable, a larger time granularity is used to reduce the amount of calculation. This optimization strategy helps to improve the utilization efficiency of system resources and reduce computing costs. At the same time, the time window adjustment coefficient calculation formula and the time granularity adjustment formula in the above technical solution have certain scalability and flexibility. By adjusting the parameters and coefficients in the formula, it can adapt to the needs of different application scenarios and improve the applicability and flexibility of the solution. By dynamically adjusting the time window granularity, the solution can provide more accurate and timely load data analysis results and provide stronger support for decision-making. This helps to improve decision-making efficiency, reduce decision-making risks, and improve overall operational benefits.

[0036] In summary, the technical effects of this technical solution in terms of performance indicators are mainly reflected in dynamic adjustment and adaptability, improved accuracy and stability, optimized resource utilization, scalability and flexibility, and improved decision-making efficiency. These effects jointly improve the overall performance and practical value of the solution.

[0037] In this embodiment, the feature recognition further includes: Acquire power data from the power system, including real-time data such as load, voltage, and current of each distribution node, perform redundancy and dimensionality reduction processing on the power data, perform feature extraction on the processed power data, and obtain an initial feature set of the power data according to the extraction result, including load features, time series features, etc.; Identify and extract key features from the initial feature set, and use each key feature in the key feature subset as a model output sample to train a preset network model to generate a load recognition model; Obtaining target power consumption characteristics corresponding to target time series data of each key feature in the key feature subset based on the load identification model; The change of the target power consumption characteristic of each load in the power data is obtained, and the power consumption change rule of each load is determined according to the change.

[0038] In this embodiment, a unified feature matrix is ​​constructed through feature-level fusion, combined with feature recognition, feature screening and dynamic adjustment of the time series window, which effectively improves the accuracy and adaptability of load forecasting. Feature recognition not only calculates the correlation coefficient between the feature and the power load, but also obtains real-time data from the power system and establishes an initial feature set. This process further refines the identification of key features. The load quantity identification model obtains the target power consumption characteristics corresponding to the target time series data of the key features, ensuring that the power consumption change rules of each load quantity can be accurately determined, thereby improving the model's ability to capture the trend of power load changes, significantly enhancing the adaptability and prediction accuracy of the load forecasting model to complex power system conditions, and providing strong technical support for the stable operation and resource optimization of the power system.

[0039] In this embodiment, the prediction result of the future load in step 3 includes: Obtain the prediction accuracy of each key feature as the first indicator for evaluating model performance; Calculate the dynamic change characteristic value of each key feature (such as load fluctuation amplitude and trend stability) as the second indicator for model performance evaluation; Count the frequency and data coverage of each feature in the data set as a supplementary performance indicator; According to the prediction accuracy, dynamic characteristic value and data coverage of each key feature, the confidence of the corresponding feature is calculated and converted into a weight value; Based on the weight value, the prediction results of the power data corresponding to each data tag are weighted averaged to obtain the final load forecast result, and the importance of each key feature in the forecast result is quantified; Based on the weight value of each key feature, the prediction results of each key feature output by the load forecasting model are weighted averaged, and the final future load forecast result is generated based on the weighted calculation result, thereby improving the overall accuracy and reliability of the forecast result.

[0040] In this embodiment, a comprehensive model evaluation system is formed by introducing prediction accuracy, dynamic characteristic values ​​and data coverage, dynamically allocating weights according to the importance of each key feature, enhancing the model's adaptability to complex data, effectively integrating the prediction results of different features, and reducing the impact of single feature errors on the overall prediction.

[0041] In this embodiment, the step 3 generates the final future load forecast result, and also includes determining the change pattern of the power load and identifying the power consumption pattern. The specific steps are as follows: Analysis of electricity consumption change patterns: Based on the electricity consumption variation rules of each load, identify and classify the electricity consumption characteristics during peak, valley and off-peak periods, and analyze the load fluctuation patterns during different periods; Combine meteorological data to determine the changing pattern of load affected by meteorological factors, identify the causes of seasonal peaks and troughs, and combine holiday information; Analyze the load fluctuation characteristics during holidays based on holiday information, distinguish the power consumption patterns during normal working days and holidays, and identify the load change patterns during special periods; Economic activity correlation analysis: Based on the correlation between economic indicator data such as GDP, industrial added value, and residents' consumption level and the electricity consumption change rules of load, a correlation model between load changes and economic indicator data is established to predict the long-term growth trend of load; Comprehensive correlation analysis: Based on the fused multidimensional data of time, season, economic activities, etc., a multidimensional correlation model of power load changes is constructed to obtain the coupling relationship between load changes and multidimensional data; The identified patterns of electricity consumption changes are verified based on the trained load forecasting model to ensure the accuracy and applicability of the patterns, and the electricity consumption pattern is optimized by combining multi-source data obtained in real time.

[0042] In this embodiment, the electricity consumption change rules are integrated into the load forecasting model to enhance the model's adaptability to complex scenarios and help the model capture load change trends more accurately. By identifying the electricity consumption change rules for each load amount, the change patterns of power load over time, season, economic activities and other factors are determined, the electricity consumption patterns during peak and trough periods are identified, and the long-term change trends of load with economic growth and industrial structure adjustment are monitored.

[0043] In this embodiment, the step 5 of identifying the abnormal type of the prediction result specifically includes: Classify and mark the emergency information in the emergency database, and associate the mark with the timestamp, impact area and load change data of the corresponding emergency information; Compare the residuals between the actual load data and the predicted results, identify abnormal data points that exceed the normal fluctuation range, match the detected abnormal data points with the emergency event information, and identify the abnormal type based on the matching results; Abnormal types include: meteorological abnormalities: caused by extreme weather events, such as high temperatures, snowstorms, etc.; equipment failure abnormalities: caused by power grid equipment failure, maintenance or accidents; economic activity abnormalities: caused by sudden economic policies, sudden increases or decreases in industrial activities; holiday abnormalities: load changes caused by holidays or major social events; Re-evaluate the forecast results after exception handling to ensure that the revised load forecast is consistent with the actual situation, identify the load change pattern before and after the event, and verify the effect of exception handling; A feedback mechanism between exception handling and load forecasting results is established, and the load forecasting model is adjusted based on real-time forecast accuracy feedback data.

[0044] In this embodiment, through refined data preprocessing, in-depth abnormal type identification and matching, and the establishment of an effective feedback mechanism, the accuracy and robustness of load forecasting are significantly improved, and the load change trend can be captured and accurately analyzed in a timely manner, and the impact of various emergencies on the power system load can be effectively identified and responded to, ensuring the stability and reliability of power supply. At the same time, it provides strong data support for the operation and scheduling, optimal resource allocation and risk management of the power system, thereby improving the overall operational efficiency and economy of the power system and meeting the demand for high-quality power supply services.

[0045] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A power system load forecasting method based on multivariate data fusion, characterized in that: The following steps are involved: Step 1: Data preprocessing: Collect multi-source data in real time based on the multi-source acquisition terminal, clean and normalize the acquired multi-source data, and perform feature coding based on the multi-source data category; Step 2: Multi-source data fusion: Fusion and feature optimization of pre-processed multi-source data, screening out key features that are highly relevant to load forecasting, and dynamically adjusting the timing window based on short-term and long-term load forecasting requirements; Dynamic adjustment of timing window, including: Extract the change rate corresponding to the current load data; Comparing the change rate corresponding to the load data with a preset change rate threshold; When the change rate corresponding to the load data exceeds a preset change rate threshold, the time granularity corresponding to the current time series window is retrieved; Taking the time granularity corresponding to the duration of the current timing window as the unit reference time; Extract load data in each unit reference time; Obtaining a load factor corresponding to each unit reference time by using the load data in each unit reference time; The load factor corresponding to each unit reference time is obtained by the following formula: ； Among them, U represents the load factor corresponding to each unit reference time; n represents the number of load data collections corresponding to each unit reference time; P represents the corresponding change rate value when exceeding the preset change rate threshold; W i represents the load data value corresponding to the i-th load data collection; W i+1 represents the load data value corresponding to the i+1th load data collection; W z represents the middle value of the load data corresponding to n load data collections; P z It represents the change rate between the load data corresponding to each unit reference time and the load data corresponding to the previous unit reference time; Retrieve the middle value of the load factor from the load factor corresponding to each unit reference time; comparing the load factor intermediate value with a preset load factor threshold; When the load factor intermediate value exceeds a preset load factor threshold, the timing window is dynamically adjusted; Step 3: Multi-model load forecasting: Build and train the corresponding load forecasting model based on multi-source data categories, and use the trained load forecasting model to perform load forecasting to obtain the forecast results of future loads; Step 4: Forecast result evaluation: Evaluate the model accuracy through model evaluation indicators, adjust the structure and parameters of the load forecast model based on the evaluation results and combined with real-time forecast accuracy feedback data, and visualize the final future load forecast results; Step 5: Abnormal response: Establish an emergency event database, identify the abnormal type of the prediction result based on the emergency event database, and perform abnormal processing response according to the abnormal type.

2. The method for power system load forecasting based on multivariate data fusion according to claim 1, characterized in that: The real-time collection of multi-source data based on the multi-source collection terminal specifically includes: Determine the data source: collect historical load data, meteorological data, holiday information and economic indicator data of the power grid in real time; Data preprocessing: remove duplicates and irrelevant information from the collected multi-source data, and standardize and convert multi-source data of different dimensions; Feature encoding processing: Perform one-hot encoding and label encoding processing on the category features of multi-source data. Among them, one-hot encoding is used to convert the category features of holiday information and meteorological data into binary vectors, and label encoding is used to process the categorical variables of historical load data; Consistency check: Align the timestamps and sampling frequencies carried by multi-source data, check and unify the measurement units and scales of multi-source data, and generate continuous and synchronized time series data.

3. The method for power system load forecasting based on multivariate data fusion according to claim 2, characterized in that: Step 2: Multi-source data fusion, specifically: Multi-source data fusion: After consistency verification, multi-source data is fused at the feature level to build a unified feature matrix; Feature identification: Calculate the correlation coefficient between each feature in the feature matrix and the power load, and identify the key features related to the power load; Feature screening: Calculate the importance weight of each key feature and recursively remove the key features with the lowest weight until the preset number of features is reached to determine the final key feature subset; Dynamic adjustment of time window: Dynamically adjust the time window size according to the load forecasting requirements of different time granularities, determine the changing characteristics of load data at different time lengths, and use the sliding window method to optimize the window length based on the forecast accuracy feedback data.

4. The method for power system load forecasting based on multivariate data fusion according to claim 3, characterized in that: Dynamically adjust the timing window, including: When the load factor intermediate value exceeds the preset load factor threshold, the maximum time granularity value and the minimum time granularity value corresponding to the historical load data are retrieved; Taking the duration corresponding to the maximum value of the time granularity as the first unit reference time; Taking the duration corresponding to the minimum value of the time granularity as the second unit reference time; Under the first unit reference time, obtaining a load factor corresponding to each first unit reference time as a first load factor data set; Under the second unit reference time, obtaining a load factor corresponding to each second unit reference time as a second load factor data set; Obtaining a time window adjustment coefficient using the load factors included in the first load factor data set and the second load factor data set combined with an intermediate value of the load factors corresponding to the current unit reference time; The time window adjustment coefficient is obtained by the following formula: ； Wherein, K represents the time window adjustment coefficient; m represents the number of load factors included in the first load factor data set; k represents the number of load factors included in the second load factor data set; U z Indicates the middle value of the load factor corresponding to the current unit reference time; U 01i represents the value corresponding to the i-th load factor in the first load factor data set; U 02i represents the value corresponding to the i-th load factor in the second load factor data set; U 01z represents the middle value of the coefficient corresponding to the m load factors in the first load factor data set; U 02z represents the middle value of the coefficient corresponding to the k load factors in the second load factor data set; T 01 and T 02 Respectively represent the corresponding durations of the first unit reference time and the second unit reference time; Using the time window adjustment coefficient to adjust the time granularity of the time window, and obtain the adjusted time granularity; The corresponding duration of the adjusted time granularity is obtained by the following formula: ； Among them, T t It indicates the duration corresponding to the time granularity after adjustment; T0 indicates the duration corresponding to the time granularity before adjustment; K indicates the time window adjustment coefficient.

5. The method for power system load forecasting based on multivariate data fusion according to claim 3, characterized in that: The feature recognition further includes: Acquire power data from the power system, perform redundancy and dimensionality reduction processing on the power data, perform feature extraction on the processed power data, and obtain an initial feature set of the power data according to the extraction result; Identify and extract key features from the initial feature set, and use each key feature in the key feature subset as a model output sample to train a preset network model to generate a load recognition model; Obtaining target power consumption characteristics corresponding to target time series data of each key feature in the key feature subset based on the load identification model; The change of the target power consumption characteristic of each load in the power data is obtained, and the power consumption change rule of each load is determined according to the change.

6. The method for power system load forecasting based on multivariate data fusion according to claim 3, characterized in that: The prediction result of the future load in step 3 includes: Obtain the prediction accuracy of each key feature as the first indicator for evaluating model performance; Calculate the dynamic change characteristic value of each key feature as the second indicator for model performance evaluation; Count the frequency and data coverage of each key feature in the data set as a supplementary performance indicator; According to the prediction accuracy, dynamic characteristic value and data coverage of each key feature, the confidence of the corresponding feature is calculated and converted into a weight value; Based on the weight value, the prediction results of the power data corresponding to each data tag are weighted averaged to obtain the final load forecast result, and the importance of each key feature in the forecast result is quantified; Based on the weight value of each key feature, the prediction results of each key feature output by the load forecasting model are weighted averaged, and the final future load forecast result is generated based on the weighted calculation result.

7. The method for power system load forecasting based on multivariate data fusion according to claim 6, characterized in that: The step 3 generates the final future load forecast result, and also includes determining the change pattern of the power load and identifying the power consumption pattern. The specific steps are as follows: Analysis of electricity consumption change patterns: Based on the electricity consumption variation rules of each load, identify and classify the electricity consumption characteristics during peak, valley and off-peak periods, and analyze the load fluctuation patterns during different periods; Combine meteorological data to determine the changing pattern of load affected by meteorological factors, identify the causes of seasonal peaks and troughs, and combine holiday information; Analyze the load fluctuation characteristics during holidays based on holiday information, distinguish the power consumption patterns during normal working days and holidays, and identify the load change patterns during special periods; Economic activity correlation analysis: Based on the correlation between economic indicator data evaluation and load power consumption change rules, a correlation model between load change and economic indicator data is established to predict the long-term growth trend of load; Comprehensive correlation analysis: Based on the fused multidimensional data, a multidimensional correlation model of power load changes is constructed to obtain the coupling relationship between load changes and multidimensional data; The identified patterns of electricity consumption changes are verified based on the trained load forecasting model, and the electricity consumption pattern is optimized by combining multi-source data acquired in real time.

8. The method for power system load forecasting based on multivariate data fusion according to claim 7, characterized in that: The model evaluation indicators in step 4 specifically include: mean absolute error indicator, mean square error indicator, root mean square error indicator, and mean absolute percentage error indicator.

9. The method for power system load forecasting based on multivariate data fusion according to claim 8, characterized in that: The abnormal type of the prediction result identified in step 5 specifically includes: Classify and mark the emergency information in the emergency database, and associate the mark with the timestamp, impact area and load change data of the corresponding emergency information; Compare the residuals between the actual load data and the predicted results, identify abnormal data points that exceed the normal fluctuation range, match the detected abnormal data points with the emergency event information, and identify the abnormal type based on the matching results; Re-evaluate the prediction results after exception handling, identify the change pattern of load before and after the event, and verify the effect of exception handling; A feedback mechanism between exception handling and load forecasting results is established, and real-time forecast accuracy feedback data is obtained based on the feedback mechanism.

Images